In this post, we’ll take a look at the highlights for developers across Salesforce Headless 360, Lightning Web Components (LWC), Apex, Agentforce, Data 360, and Agentforce 360 Platform developer tools.

Salesforce Headless 360

Headless 360 makes every major Salesforce capability available as an API, Model Context Protocol (MCP) tool, or CLI command — accessible to any authenticated caller, whether that’s an app, a human, or an autonomous AI agent. It’s the biggest theme of the Summer ’26 release. Headless 360 brings several innovations: hosted MCP servers, new MCP tools, coding skills, and CLI and API enhancements. Let’s dive into the primary developer features related to Headless 360 available in this release.

Salesforce Hosted MCP Servers

Salesforce Hosted MCP Servers let you connect any MCP-compatible AI client, such as Claude, ChatGPT, Cursor, or custom agents, to your Salesforce org and data through the open MCP standard. Every connection uses standard OAuth authentication, so your agents interact with Salesforce data and automation in a secure, governed way. Because Salesforce hosts them, there’s no additional infrastructure to manage. 

You get two flavors: pre-built standard servers, and custom servers that you define yourself.

Standard MCP Servers

Salesforce Hosted Standard MCP Servers are now generally available. Salesforce provides several pre-built standard hosted MCP servers, including:

• SObject Servers: SObject CRUD, SOQL queries, search
• Data 360: Data 360 queries and graph traversal
• Tableau: Analytics and visualization

Custom MCP Servers

When standard MCP servers aren’t enough, you can build custom MCP servers with granular control over which tools and prompts you expose. Custom MCP servers respect the full sharing and security model you have configured for your Salesforce org. Custom MCP tools can be built from:

• Apex Actions: Expose @InvocableMethod (see docs) annotated methods as MCP tools
• Lightning Flows: Expose autolaunched flows as MCP tools
• Apex REST: Expose custom Apex REST endpoints as MCP tools
• AuraEnabled: Expose @AuraEnabled annotated Apex methods as MCP tools
• Named Query API: Expose parameterized SOQL queries as MCP tools
• Prompt Builder: Expose prompts from Prompt Builder as MCP prompts
• Agentforce: Expose Agentforce agents as MCP tools
• API Catalog: Map Salesforce API Catalog (curated registry of REST API endpoints) to MCP tools

For a complete walkthrough, including OAuth configuration details and connecting Claude Desktop and Claude Code, read Connect Claude with Salesforce Hosted MCP Servers and Expose Custom Apex as a Hosted MCP Tool for Agents.

Developer and designer MCP servers and tools

Developers and designers get a productivity boost from MCP-powered tools that bring AI assistance directly into the IDE and coding agents.

Salesforce DX MCP server (Beta):  Two important tools land here. SLDS Guideline tools speeds up UI work with instant Salesforce Lightning Design System (SLDS) styling-hook and component-blueprint guidance. ApexGuru  brings Apex code review into your coding agent from your org’s runtime metrics.  It flags and fixes anti-patterns inline, such as SOQL or DML inside loops and redundant SOQL, and its Test Case Insights surface inefficient tests that drag down coverage.

Metadata API Context MCP Server (Beta): This server now ships five granular tools instead of one. These tools provide contextual information about Salesforce metadata types to help generate accurate metadata files, with faster responses and more efficient token usage.

Data 360 MCP Server (Developer Preview):  This open-source server connects your coding agent to Data 360. Instead of exposing roughly 200 REST operations one by one, it fronts them with three facade tools — search (find a capability by intent), payload_examples (fetch a working request body), and execute (run it) — which keep the coding agent from blowing its context window. 

Omnistudio MCP Server (Beta):  This server bridges agentic and low-code development. Use your coding agent to turn requirements — plain text, screenshots, or UX mockups — into functional Flexcard templates.

B2C DX MCP Server: Modify your Storefront Next components quickly with the Figma-to-Component tool set, converting design files directly into components.

Marketing Cloud Engagement MCP Server: Securely connect external AI agents to Marketing Cloud Engagement and expose core developer capabilities like data extensions and journeys as natural-language tools.

Agent Skills for coding agents

Agent Skills are a lightweight, open format for extending AI agent capabilities with specialized knowledge and workflows. In this release, we are open-sourcing a library of Salesforce development skills. The skills are optimized to work with the Salesforce coding agent Agentforce Vibes, and are also compatible with any other coding agent, such as Claude Code or Codex.

Skills come pre-packaged with Agentforce Vibes. For any other coding agent, install them using the npx command: npx skills add forcedotcom/sf-skills.

Salesforce CLI updates

Salesforce CLI’s 220+ commands are a core part of Salesforce Headless 360. The CLI keeps shipping every week, and recent releases shipped several updates worth knowing about. We’ll look at the highlights for developers, organized by what they help you build. The emphasis is on Agentforce DX tooling and a major credential security overhaul.

Build agents from a working starting point:

• Agent project scaffolding: Spin up a runnable Agentforce sample instead of starting from scratch. The agent template generates a Local Info Agent demonstrating Apex, Prompt Template, and Flow subagents.

sf template generate project --name my-agent --template agent

• One-command agent user: Automate the setup of service agent users, eliminating the need for manual provisioning.

sf org create agent-user --first-name Service --last-name Agent --target-org my-org

Test, preview, and debug agents:

• Agent preview is GA: You can now script interactive test sessions end to end with agent preview start, send, sessions, and end.
• Trace files for diagnosis: Inspect and manage the traces recorded during a preview session to see exactly how your agent routed and acted.

sf agent trace read --session-id  --format detail --dimension actions

• Richer evaluations (Beta): Run YAML- or JSON-defined evaluation tests for higher-quality, repeatable agent testing.

sf agent test run-eval --spec specs/my-agent-testSpec.yaml --target-org my-org

Keep credentials out of your logs:

• Secrets redacted by default: Access tokens, SFDX auth URLs, and passwords are now stripped from commands like org display and org list --json, preventing accidental leaks in continuous integration (CI) and logs.
• Deliberate secret retrieval: When you actually need a credential, ask for it explicitly.

sf org auth show-access-token --target-org my-org

As always, the deep-dive details for every command and flag live in the Salesforce CLI release notes. Read them to go further.

Salesforce Platform API updates

The platform’s APIs are a big part of Headless 360. Summer ’26 ships API version 67.0, and a few changes are worth knowing about as you build.

Plan now: SOAP login() is being retired

This is the most impactful change in this release for integration owners. SOAP login() in API versions 31.0–64.0 retires in Summer ’27. Any integration that authenticates with a username and password over SOAP will stop working. Move those integrations to OAuth — using external client apps with JWT tokens — well ahead of the cutoff. A new Any API Auth user permission lets you control who can authenticate via SOAP login(), and it’s enforced by default in newly created orgs.

Chain dependent records in one GraphQL request

GraphQL mutations can now reference any field returned by an earlier operation in the same request, not just its record ID. Use @{ref.Record.FieldName.value} for a field value, @{ref.Record.Id} (or shorthand @{ref}) for the ID. This lets you create linked records in a single round trip. Below is an example body for chaining dependent request in one GraphQL

mutation CreateChain {
  uiapi(input: { allOrNone: false }) {
    AccountCreate(input: { Account: { Name: "Headless 360 Test Co" } }) {
      Record { Id  Name { value } }
    }
    ContactCreate(input: { Contact: {
      LastName: "@{AccountCreate.Record.Name.value}"
      AccountId: "@{AccountCreate}"   # == AccountCreate.Record.Id
    }}) { Record { Id  LastName { value } } }
    OpportunityCreate(input: { Opportunity: {
      Name: "@{ContactCreate.Record.LastName.value}"
      AccountId: "@{AccountCreate.Record.Id}"
      StageName: "Value Proposition"
      CloseDate: "2026-12-31"
    }}) { Record { Id } }
  }
}

You can use a Beta Salesforce CLI command to execute any GraphQL as shown below.

sf api request graphql --body  --target-org my-org

Use JWT access tokens with SOAP API

SOAP API now accepts JWT-based access tokens from Salesforce OAuth flows in the sessionId header element, reaching parity with REST authentication and making token sharing with external services safer.

Connect REST API rate limits are relaxed

Orgs have been migrated off the restrictive per-user, per-application, per-hour Connect REST API limit and onto the more generous per-org, per-24-hour Salesforce Platform API limit — the same pool that every other API call shares. Only requests that require Chatter keep the old hourly throttle. The identical change applies to Connect in Apex.

User Interface API CSRF token

A new GET /ui-api/session/csrf resource (see docs) returns a token that you can use to protect User Interface API requests from third-party forgeries.

LWC updates

Summer ’26 is a maturity release for LWC. The big themes: cleaner architecture (state finally lives outside your components), a quicker edit-and-preview cycle (real-time previews in the browser and your IDE), better defaults (more virtualization, tighter security), and a new bridge to Agentforce. 

Here are the five features most likely to change how you build — each with the problem it solves.

• State Managers (GA): These pull data and its logic out of components into a reusable, testable layer.
• API 67.0 niceties: This includes zero-JS accordions via grouped <details>, plus faster hot reload.
• Secure downloads: LWS now blocks data: URIs, so generate files the right way.
• Dynamic lists (in Developer Preview): This renders thousands of rows smoothly with built-in virtualization.
• lightning/accApi: This new module lets your components open and drive the Agentforce panel.

State Managers for LWC

State Managers is the most consequential LWC feature going GA during this release. State managers move data and the logic that mutates it out of components into a dedicated layer. Build one as a plain JS module with the new defineState primitive from @lwc/state, which gives you three building blocks:

• atom(value): Reactive state, read through .value
• computed([deps], fn): A derived value that recomputes when a dependency changes
• setAtom(atom, value): The only way to update an atom

defineState returns a factory — each call yields a fresh, independent instance. The essential shape: one atom as the source of truth, a derived computed, and an action that mutates via setAtom.

Below is example code that demonstrates a state manager in action, handling a cart:

cartStateManager.js

import { defineState } from '@lwc/state';
export default defineState(({ atom, computed, setAtom }) => {
    const items = atom([]);                       // source of truth
    const count = computed([items], (l) => l.length); // derived
    const addItem = (item) =>                       // action
        setAtom(items, [...items.value, item]);
    return { items, count, addItem };
});

A component imports the manager module, calls the factory, then reads reactive state through .value — in JS or the template, which re-renders automatically on change. No data logic, no manual subscription:

cartSummary.js

import { LightningElement } from 'lwc';
import createCartStateManager from 'c/cartStateManager';
export default class CartSummary extends LightningElement {
    cart = createCartStateManager();
}

cartSummary.html

<template><h2>Your Cart ({cart.value.count})</h2><p>
</p></template>

For complete, runnable examples, see the open-source lwc-recipes repo. Because each instance is isolated, managers are trivially unit-testable.

Salesforce also ships built-in Lightning state managers that wrap Lightning Data Service (LDS) access to the most common UI API data and metadata — for records, object info, page layouts, and related lists (for example, lightning/stateManagerRecord and lightning/stateManagerObjectInfo). They follow the same pattern as the ones you write and participate fully in LDS caching, normalization, and subscriptions, so reach for them before rolling your own.

LWC API version 67.0: Group <details> and faster HMR

To enable the features of this release, update your bundle .js-meta.xml by setting <apiVersion>67.0</apiVersion>.

Two wins by using the 67.0 API version: hot module reloading (HMR) is faster and more memory-efficient, and you can now group native <details> elements with the name attribute for a zero-JavaScript exclusive accordion (it threw a compiler warning before 67.0). Same name = only one open at a time.

faqAccordion.html

<details key={faq.id} name="summer26-faq">
    <summary>{faq.question}</summary>
    <p>{faq.answer}</p></details>

Secure file downloads: LWS blocks data: URIs

Lightning Web Security (LWS) adds a batch of API distortions in this release. The one most likely to break code: HTMLAnchorElement.prototype.href now blocks the data: URI scheme. If you trigger client-side downloads by setting an anchor’s href to a data: URL, that stops working.

The fix is the supported pattern anyway — build a blob in JavaScript and use a blob: object URL (origin-bound, and revoked after use).

secureDownload.js

// LWS-safe: blob: URL, not data:
const blob = new Blob([this.content], { type: 'text/plain' });
const url = URL.createObjectURL(blob);
anchor.href = url;
anchor.download = this.fileName;
anchor.click();
URL.revokeObjectURL(url); // release memory

Other new distortions include Element.getAttribute, innerHTML/outerHTML getters, MutationObserver.observe, the IndexedDB factory, Promise.then/catch/finally, and more — with matching ESLint rules. Run the updated ESLint package and review the LWS Distortion Viewer before you upgrade the components to this release.

Load large lists dynamically (Developer Preview)

Rendering thousands of rows used to choke the browser, the new lightning-dynamic-list-container (see docs) and lightning-dynamic-list-item (see docs) base components use virtualization. They render only the rows in the viewport and stream the rest in as you scroll, from 50 items to 5,000.

dynamicContactList.html

<!-- Container needs a bounded height for scroll + virtualization -->
<div style="height: 400px">
  <lightning-dynamic-list-container
      onrenderlistitems={handleRenderListItems}
      onloadmore={handleLoadMore}>
    <template for:each={visibleContacts} for:item="contact">
      <lightning-dynamic-list-item
          key={contact.id} item-id={contact.id}>
        <div>{contact.name}</div>
      </lightning-dynamic-list-item>
    </template>
  </lightning-dynamic-list-container>
</div>

The container fires renderlistitems as you scroll (which slice to render) and loadmore near the end:

dynamicContactList.js

handleRenderListItems(event) {
    const { startIndex, endIndex } = event.detail;
    this.visibleContacts = this.allContacts.slice(startIndex, endIndex);
}

You also get focus preservation and built-in accessibility (screen readers, Home/End/Arrow nav, Browse-Mode hint). 

Here are some recommendations for working with dynamic lists: keep container and item adjacent, give every item a unique item-id, and don’t set overflow: scroll on your own container — the component handles scrolling.

Talk to Agentforce from LWC with new LWC module lightning/accApi

The standout new module is lightning/accApi (see docs) — the Agentforce Conversation Client API. This headless module lets your LWC components drive the native Agentforce side panel in Lightning Experience: open it, close it, point it at a specific Employee Agent, and send natural-language utterances. Think of a “Summarize this record” button, or a context-aware launcher in a console sidebar.

The entire API is three async methods:

All three return a Promise and are queued, running in sequence. Note execute does not return the reply — the conversation renders in the panel, not your component. Import the methods and call them.

agentforceLauncher.js

import { open, close, execute } from 'lightning/accApi';

@api botId; // design-time property, set on the page

async handleOpen() {
    await open(this.botId);
}
async handleQuickAction(event) {
    const { utterance } = event.currentTarget.dataset;
    await execute(utterance, this.botId); // wrap in try/catch + toast
}

// Deploying this class at API 67.0 fails with:
//   "WITH SECURITY_ENFORCED is no longer supported, use WITH USER_MODE instead."
[SELECT Id FROM Account WITH SECURITY_ENFORCED LIMIT 1];

try {
    List a = [SELECT Id, Name, AnnualRevenue FROM Account WITH USER_MODE LIMIT 1];
} catch (QueryException e) {
    // returns Map fieldNames>
    Map<String, Set> blocked = e.getInaccessibleFields();
}

String payload = '''
{
    "Account": "${accountName}",
    "Last Updated": "${date}"
}'''.template(new Map<String, Object>{
    'accountName' => 'My Account',
    'date' => Datetime.newInstance(2018, 11, 15, 8, 0, 0)
});

@IntegrationTest
public with sharing class MyServiceIntegrationTest {
    @IntegrationTest
    public static void testServiceInteraction() {
        Account a = new Account(Name = 'Integration Test Account');
        insert as user a;
        IntegrationTest.commitTestOnly();           // data survives mid-test
        Account r = [SELECT Name FROM Account WHERE Id = :a.Id WITH USER_MODE];
        Assert.areEqual('Integration Test Account', r.Name);
    }
    @TearDown
    public static void tearDown() {
        delete as user [SELECT Id FROM Account WHERE Name = 'Integration Test Account' WITH USER_MODE];
    }
}

sf api request rest "/services/data/v67.0/einstein/data-libraries" \
  --method POST --target-org my-org \
  --body '{
    "masterLabel": "Product Docs Library",
    "developerName": "Product_Docs_Library",
    "groundingSource": { "sourceType": "SFDRIVE" }
  }'
# → { "libraryId": "1JD...", "status": "IN_PROGRESS" }

knowledge:
    rag_feature_config_id: "ARFPC_1JDcf0000024ZZ7GAM"
    citations_enabled: True

import { AgentforceService } from 'react-native-agentforce';
// Option A — Service Agent (anonymous, customer-facing)
await AgentforceService.configure({
    type: 'service',
    serviceApiURL: 'https://service.salesforce.com',
    organizationId: '00DWs00000Ip47F',
    esDeveloperName: 'Order_Support_Agent',        // the agent's API name
    serviceUISettings: { enableLightningType: true }  // render Custom Lightning Types as cards
});

// Option B — Employee Agent (internal, authenticated)
await AgentforceService.configure({
    type: 'employee',
    instanceUrl: 'https://your-domain.my.salesforce.com',
    organizationId: '00DWs00000Ip47F',
    userId: '005xx0000001234',
    agentId: '0XxWs000001DTDJK'                       // Bot Id from publishing the agent
});

await AgentforceService.setAdditionalContext({
    variables: [{ name: 'userId', type: 'Text', value: '005xx0000001234' }]
});

await AgentforceService.launchConversation();

SELECT Id, EmailOptIn__c
FROM ContactDLO__dlm
WHERE EmailOptIn__c = ''
SET OPTIONS (dataspace = 'default', honorEmptyStrings = true)

The post The Salesforce Developer’s Guide to the Summer ’26 Release appeared first on Salesforce Developers Blog.

This is the first post in a series on common anti-patterns in Lightning Web Components (LWC). In this post, you’ll learn how the three security layers in the Lightning platform work, which APIs are namespaced versus blocked, and how to avoid the most common mistakes that lead to silent failures in production. The patterns covered here apply to any custom LWC running on the Salesforce Platform.

Note: LWS distortions can change between Salesforce releases as the platform evolves. The LWS Distortion Viewer is the live source of truth for the exact behavior of every distorted API. If a pattern in this post no longer matches what you observe in your org, check the Distortion Viewer first before filing a bug.

Three security layers instead of one

Your LWC code runs inside three independent security layers. Knowing which layer does what saves you hours of debugging.

Lightning Web Security (LWS) is the namespace isolation layer. It places your component’s JavaScript in a sandboxed environment and applies distortions to browser APIs. Distortions don’t simply block APIs — they namespace storage, sanitize HTML on shared DOM elements, sandbox code evaluation, and block only a small number of APIs that could escape the sandbox. You can see every distortion and its exact behavior in the LWS Distortion Viewer.

The LWC framework enforces shadow DOM scoping, prevents access to legacy Aura globals like $A, and manages the component lifecycle.

Content Security Policy (CSP) and platform-level restrictions block inline scripts, external CDN loading, and certain URL schemes. These operate independently of LWS.

When something doesn’t work, your first question should be: which layer is responsible? Getting this wrong leads to workarounds that solve the wrong problem.

Five key anti-patterns

The patterns below fall into five categories. The first three describe how LWS distorts APIs — namespacing, sanitizing/sandboxing, and outright blocking. The fourth covers what happens at the boundary between namespaces. The fifth covers patterns that look like LWS issues, but actually come from the LWC framework or CSP. Identifying the right category is usually the fastest path to the right fix.

1. Namespaced APIs

LWS does not block these APIs — it applies a namespace isolation layer. The APIs function as expected within your own component set, but they remain blind to data established by the platform or by components residing in other namespaces.

Assuming localStorage and sessionStorage are global

LWS does not block localStorage (see docs) or sessionStorage (see docs). It namespaces them. Every key is transparently prefixed with LSKey[namespace], so each namespace gets its own isolated storage. When you call localStorage.setItem('myKey', 'value') in namespace c, Salesforce actually stores LSKey[c]myKey. Your component sees only its own keys.

This means localStorage works — but only within your namespace. If you expect to read keys set by the platform or by a component in a different namespace, you’ll get back null with no error.

// Returns null — the key was set outside this namespace's sandbox
connectedCallback() {
    const token = window.localStorage.getItem('sessionToken');
}

Use localStorage for same-namespace persistence like caching user preferences. For cross-namespace data or system-level state, use custom settings, Apex calls, or the Lightning Message Service.

Assuming document.cookie is global

Like storage, document.cookie (see docs) is namespaced by LWS, not blocked. The getter returns only cookies belonging to your sandbox, and the setter adds a sandbox prefix to new cookie keys. Platform cookies and cookies from other namespaces are invisible.

Don’t rely on document.cookie to read platform or cross-namespace cookies. Use server-side Apex to access session or authentication state instead.

2. Sanitized and sandboxed APIs

LWS does not block APIs like innerHTML (see docs), eval() (see docs), or Function() (see docs). Instead, it applies specific distortions based on the execution context. The key to avoiding silent failures is understanding exactly how the platform modifies these behaviors rather than attempting to bypass them entirely.

DOM mutation on shared elements

LWS runs in the main window, where <html>, <head>, and <body> are shared across all components. LWS protects these shared elements by sanitizing HTML strings and restricting which child elements you can add. For elements inside your component’s own shadow DOM, these APIs work normally.

Here’s the key distinction: innerHTML, outerHTML, insertAdjacentHTML, and related APIs are not blocked. They’re sanitized when targeting shared elements, and unrestricted on elements your component owns.

// innerHTML works on component-owned elements — but this is still an XSS risk
renderedCallback() {
    const container = this.template.querySelector('.container');
    container.innerHTML = `${this.userProvidedValue}`;
}

LWS will sanitize this if container is a shared element, but the real issue is the Cross Site Scripting (XSS) risk. Relying on LWS sanitization as a security mechanism is fragile — it protects shared DOM elements, not your component’s shadow DOM.

Use LWC’s declarative template directives (lwc:if, for:each, template expressions) for dynamic content. For trusted rich text from a CMS, use lightning-formatted-rich-text (see docs).

Code evaluation is sandboxed, not blocked

This one surprises many developers: eval() (see docs) and the Function() (see docs) constructor are not blocked by LWS. They run inside the sandbox — LWS ensures the evaluated code executes in the same sandboxed context as your component. Passing a string to setTimeout() (see docs) or setInterval() (see docs) works the same way.

eval('1 + 1') returns 2 in a Salesforce org. But eval() is still a bad practice. It makes code harder to analyze, blocks compiler optimizations, and creates injection vectors if the evaluated string includes user input.

// eval works in the sandbox — but don't do this
processRule(ruleString) {
    return eval(ruleString);
}

Replace dynamic code evaluation with explicit logic, a lookup table, or a strategy pattern. If you need to evaluate user-defined expressions, implement a safe parser scoped to the expression language you support.

3. APIs that are actually blocked

APIs that throw at runtime

A small number of APIs are genuinely blocked by LWS because they provide direct paths to escape the sandbox. Calling any of these throws a runtime exception.

Use LWC’s declarative template system for HTML rendering and lightning/platformResourceLoader (see docs) for script loading. There is no supported workaround for Workers inside the LWS sandbox. If your use case requires them, consider an iframe-based isolation strategy.

Network requests are not broadly restricted

fetch() (see docs), XMLHttpRequest (see docs), navigator.sendBeacon() (see docs), and fetchLater() (see docs) work normally under LWS. The only distortion blocks requests to URLs containing /aura or /webruntime — these are internal framework endpoints that your components should not access directly. All other network requests work normally, subject to standard Cross-Origin Resource Sharing (CORS) rules and your org’s CSP configuration.

// Blocked — hitting internal framework endpoints
fetch('/aura?aura.ApexAction.execute=1', { method: 'POST', body: payload });

Use Apex @AuraEnabled methods via the wire service or imperative calls instead.

4. Cross-namespace boundaries

LWS employs a proxy membrane to enforce strict isolation across namespace boundaries. Consequently, objects traversing this boundary exhibit different behaviors compared to those residing within a single namespace. For further details on the performance impacts of this membrane proxying, consult the LWS Performance documentation.

Cross-namespace object isolation

When your component receives an object from a different namespace — through an @api property, event detail, or a shared service — and you mutate that object in place, the change is not propagated back outside the sandbox. No error is thrown. The originating component just keeps seeing the original value. This is one of the hardest LWS traits to diagnose because the only symptom is stale data.

For more on the performance implications of cross-namespace membrane proxying, see the LWS Performance documentation.

// childComponent.js — namespace 'c'
@api record;

handleEdit() {
    // Mutation is silently absorbed by the LWS proxy — parent never sees it
    this.record.Name = 'Updated Name';
}

Clone the received object before modifying it, then communicate changes back with a CustomEvent.

handleEdit() {
    const updated = { ...this.record, Name: 'Updated Name' };
    this.dispatchEvent(new CustomEvent('recordchange', {
        detail: { ...updated }
    }));
}

Passing objects across namespace boundaries

Objects passed between components in different namespaces are wrapped in LWS Proxy objects (see docs). Proxy unwrapping can fail silently, giving you null or incorrect values on the receiving end. Browser extensions that listen to custom events also receive null for detail when it contains a proxied object.

// Namespace 'foo' — dispatching
this.dispatchEvent(new CustomEvent('selected', {
    detail: { record: this.selectedRecord }  // may arrive as null in another namespace
}));

Ensuring your object can be structured cloned is the key. Or, serialize the payload to JSON before dispatching and parse it on receipt. A plain string crosses the namespace boundary without proxy wrapping.

// Namespace 'foo'
this.dispatchEvent(new CustomEvent('selected', {
    detail: JSON.stringify(this.selectedRecord)
}));

// Namespace 'bar'
handleSelected(event) {
    const record = JSON.parse(event.detail);
}

5. Misattributed restrictions

These patterns fail at runtime, but the cause is not LWS. Blaming LWS leads to the wrong fix.

Referencing legacy Salesforce globals

The LWC framework (not LWS) blocks access to $A, Aura, Sfdc, and sforce in Lightning web components. These belong to the Aura framework or are deprecated platform globals. Any reference to them in LWC code fails at runtime.

This is the most common mistake when migrating Aura components to LWC.

// Pattern carried over from Aura — fails at runtime
connectedCallback() {
    const userId = $A.get('$SObjectType.CurrentUser.Id');
}

Use the LWC platform equivalents instead.

import userId from '@salesforce/user/Id';
import { getRecord } from 'lightning/uiRecordApi';
import NAME_FIELD from '@salesforce/schema/User.Name';

For a complete mapping, see the Migrate Aura Components to LWC documentation.

Loading third-party scripts from external CDNs

Loading a JavaScript library with a <script> tag pointing to an external CDN URL violates Salesforce’s CSP. The platform blocks it before LWS even comes into play. Beyond the CSP violation, external CDN loading introduces a versioning risk: the CDN may update the library at any time, silently introducing sandbox incompatibilities.

connectedCallback() {
    const script = document.createElement('script');
    script.src = 'https://cdn.example.com/somelib/v3.min.js'; // CSP violation
    document.head.appendChild(script);
}

Download the library, upload it as a Static Resource, and load it with lightning/platformResourceLoader.

import { loadScript } from 'lightning/platformResourceLoader';
import SOMELIB from '@salesforce/resourceUrl/somelib';

async renderedCallback() {
    if (this._libLoaded) return;
    this._libLoaded = true;
    await loadScript(this, SOMELIB);
    this.initializeLib();
}

Test the static-resource version in a sandbox before promoting to production. If the library requires APIs that LWS blocks (like Workers or document.write), it won’t work inside Salesforce regardless of how you load it.

Conclusion

These anti-patterns share a root cause: misunderstanding which security layer does what. LWS doesn’t block most things — it namespaces, sanitizes, and sandboxes. The APIs it actually blocks are a small, specific set. Meanwhile, the LWC framework and CSP enforce their own independent restrictions.

The mental model to remember: LWS is a namespace isolation layer, not a blanket firewall. When an API seems “blocked,” it’s usually namespaced — you’re looking at an empty namespace, not a wall. When in doubt, check the LWS Distortion Viewer for the exact behavior.

In the next post in this series, we’ll cover data and communication: the Wire Service, @api properties, and event handling. These are the three surfaces where most component-to-component bugs originate, and where a few common misunderstandings cause the most trouble.

Resources

• LWS Distortion Viewer — the definitive reference for every LWS distortion and its behavior
• Lightning Web Security architecture
• LWS Performance — performance implications of cross-namespace membrane proxying
• Lightning Web Security Developer Guide
• Migrate Aura Components to LWC

About the author

Tim Dionne is a PMTS on the Customer Centric Engineering team. He’s worked on many UI features of Salesforce over the years, starting with VisualForce, Aura Components, and Lightning Web Components with an emphasis on Lightning Web Security and Lightning Data Service.

The post Security Anti-Patterns in Lightning Web Components appeared first on Salesforce Developers Blog.

In this post, you’ll learn the basic syntax and mental model, typical use cases (including e‑commerce examples), and pilot gates and current constraints, so you can try it safely in a sandbox.

The pilot is open to any Salesforce customer — contact your Salesforce Account Executive or Customer Success Manager to request access.

Why use FORMULA() in SOQL WHERE?

When a filter depends on something that you calculate from fields, such as margin, revenue per head, days idle, or days late to ship, the usual choices are: add a formula field, maintain a duplicate rule in Apex, or query too many rows and throw away what you do not need. 

FORMULA('…') in WHERE is for the case where the rule belongs with the query: you describe the computed condition once, the platform evaluates it as part of filtering, and callers (Apex, APIs, downstream jobs) get a smaller, already qualified result set. This is especially valuable when you can not or do not want to change the data model for every subscriber org, a common scenario for ISVs or developers who want to keep operational segmentation logic visible in SOQL instead of buried in loops.

For example, say you’re working with an Order__c object and need to find every order where profit (revenue minus cost) exceeds $150. Without FORMULA(), the rule lives in Apex after a broader query; with FORMULA(), the rule moves into the WHERE clause itself.

For example, say you’re working with an Order__c object and need to find every order where profit (revenue minus cost) exceeds $150. Without FORMULA(), the rule lives in Apex after a broader query; with FORMULA(), the rule moves into the WHERE clause itself.

// Without FORMULA(): business rule lives in Apex after a wider query

 List rows = [SELECT Id, Name, Revenue__c, Cost__c FROM Order__c];
  List highMargin = new List();
  for (Order__c o : rows) {
      if (o.Revenue__c != null && o.Cost__c != null
          && (o.Revenue__c - o.Cost__c) > 150) {
          highMargin.add(o);
      }
  }

--With FORMULA(): same rule expressed in the WHERE clause 

SELECT Id, Name FROM Order__c WHERE FORMULA('Revenue__c - Cost__c') > 150

The formula expression supports addition and subtraction (+ and -). Expressions can evaluate to DOUBLE, INTEGER, DATETIME, DATE & CURRENCY.

In practice:

• INTEGER behaves like DOUBLE
• DATE behaves like DATETIME

So the rest of this post focuses on the three illustrative types: DOUBLE, DATETIME, and CURRENCY.

This pattern fits Apex and SOQL developers, ISV teams shipping logic across subscriber orgs where schema changes are hard, and anyone trying to move off the “query a large candidate set, filter in code” approach.

Beyond fitting that audience, FORMULA() in WHERE matters today because it:

• Eliminates unnecessary formula fields
• Reduces Apex complexity
• Improves performance by filtering at the query level
• Is especially valuable for ISVs who can’t modify customer schemas

Syntax: FORMULA() in WHERE

Today, FORMULA() in SOQL is available as a pilot capability: it is only supported in the WHERE clause (not HAVING).

The general shape is:

WHERE FORMULA('…') <operator> <literal>

The string is a formula expression evaluated for each row as part of filtering.

Example: E-commerce order management 

Imagine that you’re building an order analytics dashboard for a growing e-commerce platform.

Your custom Order__c object tracks:

• Revenue__c (Currency): Total order value
• Cost__c (Currency): Fulfillment cost
• OrderDate__c (Date): When customer placed order
• ShipDate__c (Date): When order shipped
• Status__c (Text): Order status

Previously, filtering orders by calculated metrics meant either:

1. Creating formula fields like Profit__c, ShippingDelay__c, ProfitMargin__c
2. Querying all orders and filtering in Apex loops

Let’s see how FORMULA() eliminates both.

The screenshot below shows a Sample Order__c records table that includes seven orders (ORD-001 through ORD-007) with Revenue, Cost, OrderDate, ShipDate, and Status columns used throughout the FORMULA() examples.

Three key use cases

For our e-commerce example, FORMULA() is particularly helpful in the following three use cases:

1. Find high-profit orders using currency arithmetic.
2. Detect late shipments using datetime arithmetic.
3. Filter by premium order criteria using combined conditions.

1. Find high-profit orders (CURRENCY arithmetic)

Our business requirement for this first use case is to identify orders with profit > $250 for priority fulfillment review.

The old way

 // Query everything, filter in memory
  List allOrders = [
      SELECT Id, Name, Revenue__c, Cost__c
      FROM Order__c
      WHERE Status__c = 'Shipped'
  ];

  List highProfitOrders = new List();
  for (Order__c order : allOrders) {
      if (order.Revenue__c != null &&
          order.Cost__c != null &&
          (order.Revenue__c - order.Cost__c) > 250) {
          highProfitOrders.add(order);
      }
  }

// Result: 3 orders (ORD-001, ORD-003, ORD-005)

The new way

SELECT Id, Name, Revenue__c, Cost__c
FROM Order__c
WHERE Status__c = 'Shipped'
AND FORMULA('Revenue__c - Cost__c') > 250

The following screenshot shows a SOQL query using the FORMULA() function to filter orders by profit: SELECT Id, Name, Revenue__c, Cost__c FROM Order__c WHERE Status__c equals “Shipped” AND FORMULA('Revenue__c - Cost__c') is greater than 250.Query results show three records: ORD-001 with 500 revenue and 200 cost, ORD-003 with 800 revenue and 300 cost, and ORD-005 with 1200 revenue and 400 cost.

2. Detect late shipments (DATETIME arithmetic)

Our business requirement here is to find orders that shipped more than three days after order date for SLA analysis. 

SELECT Id, Name, OrderDate__c, ShipDate__c
FROM Order__c
WHERE FORMULA('ShipDate__c - OrderDate__c') > 3

The screenshot below shows a SOQL query using the FORMULA() function to calculate shipping delays: SELECT Id, Name, OrderDate__c, ShipDate__c FROM Order__c WHERE FORMULA('ShipDate__c minus OrderDate__c') greater than three. Query results display two records: ORD-002 ordered April 5 and shipped April 12, and ORD-005 ordered April 15 and shipped April 25, both exceeding the three-day shipping threshold.

3. Premium order criteria (combined conditions)

Our business requirement for the third use case is to find “premium” orders: high revenue (> $600) AND fast shipping (≤ 2 days).                                                                                                           

SELECT Id, Name, Revenue__c, OrderDate__c, ShipDate__c
FROM Order__c
WHERE Revenue__c > 600
AND FORMULA('ShipDate__c - OrderDate__c') <= 2

The screenshot below shows a SOQL query combining standard field filter with the FORMULA() function: SELECT Id, Name, Revenue__c, OrderDate__c, ShipDate__c FROM Order__c WHERE Revenue__c is greater than 600 AND FORMULA('ShipDate__c minus OrderDate__c') is less than or equal to two. Query results show two records: ORD-003 with 800 revenue, ordered April 10 and shipped April 11, and ORD-007 with 650 revenue, ordered April 20 and shipped April 21, with both records meeting the high-revenue and fast-shipping criteria.

Conclusion

By applying the ideas in this post, you can express computed filters directly in SOQL. FORMULA() in WHERE turns the question from “what rows might be relevant?” into “what rows already match my computed rule?” with less schema churn. This tends to shrink Apex branching, reduce one-off formula fields for query-only rules, and make the business condition readable next to the SELECT, which helps both integrations and teams reviewing SOQL in code review and operations.

Beyond individual queries, the same pattern reinforces a broader habit: keep data access and the rule that defines “the right rows” together, so batch jobs, services, and packaged logic stay easier to maintain, especially when subscriber org schemas are not yours to reshape at will. As always, treat what you validate in a non-production org as the contract for syntax, supported types, and supported operators.

We would love your feedback on the capabilities of this pilot, and your comments will influence the GA release. Please post your questions and/or feedback on the Pilot to the Salesforce Developers Trailblazer Community. Just let us know!

Resources

• Pilot enrollment: Contact your Salesforce Account Executive or Customer Success Manager to request access to the SOQL FORMULA() pilot
• Trailhead: SOQL for Admins – Create SOQL Queries to get data from your Salesforce org

About the author

Dikshita Patel is a Software Engineer on Salesforce’s Enterprise API team, where she builds Platform APIs allowing developers, partners, and customers to query and access Salesforce data without using the Salesforce user interface. You can follow and connect with her on LinkedIn.

The post Simplify Your SOQL Queries Using SOQL FORMULA() in WHERE appeared first on Salesforce Developers Blog.

The MCE MCP Server is Salesforce’s first-party, enterprise-grade, hosted MCP server for Marketing Cloud Engagement. It’s a first step towards a truly headless experience, exposing MCE’s core marketing capabilities as tools that any MCP-compatible AI agent can call — in plain language, without writing code or navigating the UI. 

The server allows almost any external agent that supports MCP to manage data extensions, journeys, automations, and more. Note that Marketing Cloud API limits and guidelines apply (see documentation for more information).

In this post, we’ll walk you through the MCE MCP Server setup, tools, and typical use cases.

How to set up the MCE MCP Server

An agent’s permissions are a combination of the scopes in a dedicated, installed package and your own user permissions. If your installed package has access to read data extensions, but you (as a user) do not, then the MCP server does not have access. Likewise, you can limit what it can do by not granting permissions to the installed package, even if the user still has them. 

From the install package, you’ll be able to get an MCP URL that you and potentially other teammates can use. This is specific to your organization and the install package. It’s not sensitive, but it’s worth keeping in a safe place for convenience. Registering it with Claude Code, for example, is simple.

claude mcp add --transport http  https://

Other agents will be very similar. You should only need to do this once, and then authenticate from inside the agent. The details on how this works will vary depending on how the agent implements OAuth login flows, but you will be prompted to log into your Marketing Cloud account if you are not already logged in. This gives the agent a token that it can use to act on your behalf, but only for a limited time. 

For more information on MCE MCP Server setup, please see the documentation.

MCE MCP Server capabilities and tool reference

The MCE MCP Server is designed to simply wrap existing functionality provided by the Marketing Cloud Engagement APIs. For example, to get lists of data extensions, we offer a tool called sfmc_get_data_extensions. This is nearly 1:1 with a request to data/v1/customobjects on the original REST route, but made more easily discoverable for an agent. It can read the embedded documentation and discover that it needs to provide a $search variable with the search string.

As an example, this is how this tool is shown to the agent:

{
  "name": "sfmc_get_data_extensions",
  "description": "Retrieve a list of data extensions. GET /data/v1/customobjects.\n\nIMPORTANT: The $search query parameter is REQUIRED and must be a valid search term (no wildcards like * or %).\nYou must provide query_json with the structure: {\"$search\": \"term\"}\n\nExamples:\n- Search for 'customer': {\"$search\": \"customer\"}\n- Search for 'email': {\"$search\": \"email\"}\n\nSee resource 'sfmc://docs/data/v1/customobjects' for full API documentation.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query_json": {
        "type": "string",
        "description": "JSON object string containing $search parameter. REQUIRED format: {\"$search\": \"search_term\"}. The search term cannot be empty or contain wildcards (* or %)."
      }
    },
    "required": [
      "query_json"
    ]
  },
  "annotations": {
    "title": "sfmc get data extensions",
    "readOnlyHint": true,
    "destructiveHint": false,
    "idempotentHint": true,
    "openWorldHint": true
  }
}

Just as our documentation promised, the resulting tool call is simple. 

{ "$search": "test" }

The results are also pretty straightforward, even for a human to read.

Status: 200 OK
URL: https://mcfdpyw7c97kdckv65lk2j1vspd0.rest.marketingcloudapis.com/data/v1/customobjects?%24search=test

{
  "count" : 8,
  "page" : 1,
  "pageSize" : 25,
  "links" : { },
  "items" : [ {
    "id" : "bdc04e14-0209-f111-a5e8-5cba2c701e38",
    "name" : "test_7",
    "key" : "test_7",
    "description" : "Subscribers who purchased in 2018",
    "isActive" : true,
    "isSendable" : false,
    "isTestable" : false,
    "categoryId" : 1998298,
    "ownerId" : 11280369,
    "isObjectDeletable" : true,
    "isFieldAdditionAllowed" : true,
    "isFieldModificationAllowed" : true,
    "createdDate" : "2026-02-13T11:33:22.133",
    "createdById" : 11280369,
    "createdByName" : "GI Admin",
    "modifiedDate" : "2026-02-13T11:33:22.133",
    "modifiedById" : 11280369,
    "modifiedByName" : "GI Admin",
    "ownerName" : "GI Admin",
    "partnerApiObjectTypeId" : 310,
    "partnerApiObjectTypeName" : "DataExtension",
    "rowCount" : 0,
    "dataRetentionProperties" : {
      "isDeleteAtEndOfRetentionPeriod" : false,
      "isRowBasedRetention" : false,
      "isResetRetentionPeriodOnImport" : false
    },
    "fieldCount" : 15
  }, ...

It’s worth noting that the result is NOT JSON as far as the agent is concerned. It’s simply text, which in this case represents the result of making an API call to a REST endpoint. It is up to the agent to interpret this information and use it, and that’s what provides the maximum flexibility.

While the original intent of this API was to integrate with custom software, an AI agent can instead make it available interactively and with minimal friction. You don’t need to know about $search (did you forget the dollar sign?) or its syntax, the agent can deal with that. If the tool fails, it has access to the error message and relevant documentation to try again with different arguments without requiring a human to copy and paste an error code into a search engine. In this case, the tool offered sfmc://docs/data/v1/customobjects as a resource to selectively load and find out more about custom object definitions if necessary.

Another advantage is that if you are making a complex request, the AI agent (like Claude Client) can sequence tools that need to be executed in order to achieve the given functionality. For example, If you just tell the AI client to “Update the loyalty status of subscriberId 5544 to gold in XYZ DE,” the agent first gets the XYZ DE, fetches the row in DE, and updates the row without user intervention.

See the MCE MCP tool reference for more information.

What you can build with MCP + Marketing Cloud

We tested a range of use cases, from creating a new campaign journey using a single prompt to sending transactional messages in real time. AI agents are not necessarily the most creative things around, but our test agent was able to put most of the pieces we needed together, and it prompted for more context if it needed clarity on intent. Typically, you really want to hand off mundane tasks that you’d rather not deal with to an agent, such as the following.

• Create a Data Extension from a plain English description: Ask your AI agent, “Create a Data Extension for Holiday Shoppers with Email, First Name, Last Name, and Purchase Date.” The agent maps field descriptions to Marketing Cloud datatypes, sets SubscriberKey as the primary key, and returns the CustomerKey and a direct link to the new Data Extension — no Contact Builder navigation required.
• Launch a basic automated campaign without touching Journey Builder: Ask, “Create a Welcome Email journey that sends immediately when someone joins the Welcome List DE.” The agent verifies the entry source DE, builds the Entry → Email → Exit structure, resolves the email asset, and creates the journey in Draft. It then asks if you want to activate it.
• Build a nurture series conversationally: Ask, “Create a 3-email Welcome Series: send Email 1 immediately, wait 3 days, send Email 2, wait 7 days, send Email 3.” The agent parses the wait durations, names each activity descriptively, resolves all email assets, and returns the Journey Builder URL ready to review and activate.
• Add Einstein optimization to any journey step: Ask, “Add Einstein Send Time Optimization before the promotional email in my Black Friday Journey.” The agent verifies that Einstein STO is provisioned for your Business Unit, and inserts the activity immediately before the specified email.
• Propagate a new data field across every dependent DE and query in one command: Ask, “Add Propensity_Score (Number) from the Customer_Signals DE to all downstream Data Extensions and queries.” The agent traces transitive dependencies across all Query Activities and target DEs, shows you a full impact report (X DEs, Y query steps across Z automations), and — on your confirmation — adds the field to every DE, and updates every SELECT clause in dependency order. It then returns a summary of what changed and any errors encountered. This supports a dry-run mode to preview all changes before touching anything.

Having said that, we encourage you to try all your use cases with this MCP server and let us know if we are missing any crucial or nice-to-have tools on IdeaExchange.

Permissions, destructive operations, and safety guardrails

We’ve done our best to make this service as flexible and useful as possible, for as many people as possible. We decided early on to even support “destructive” operations like deleting data, even though that’s almost always a bad idea to hand off to an agent. Operations like this are annotated as destructive in a way that the agent can read, but the agent is still capable of choosing to execute them if it has the permissions to do so. It’s crucial to think through the “worst case” when assigning permissions during setup, because an agent can make mistakes. Customers are also sensitive to reputational harm if an unintentional send is triggered, so bear that in mind as well! Please make sure that you review your Installed Package scopes before using them. More details in this documentation.

Future growth

The MCE MCP Server is designed to be extended in the future, so be sure to check back often. We plan to extend coverage to other important Marketing Cloud APIs and improve what we’ve already built. Agents and the LLMs that power them are also continuously releasing and improving, so workflows that weren’t possible before might get more practical over time. 

We plan on continually improving the server and extending it to other Marketing Cloud APIs over time. Please keep an eye on our release notes, and provide feedback through your account manager or provide on IdeaExchange.

Resources

• Documentation: MCE MCP Server Setup Guide
• Documentation: MCE MCP Server Tool Reference

About the author

​​Swetha Pinninti is a Director of Engineering at Salesforce on the Marketing Cloud Einstein team. 

Patrick Frampton is a Lead Member of Technical Staff at Salesforce on the Marketing Cloud Einstein team.

The post The MCP Server for Marketing Cloud Engagement is Now GA appeared first on Salesforce Developers Blog.

This post breaks down the structure of a Salesforce agent skill using Apex as an example. For clarity, we’ll use Claude Code to highlight specific examples, but the principles apply to any agent. We’ll cover what a skill is, the project structure, and how to get started creating your first skill.

What is an agent skill?

An agent skill is a structured prompt package that transforms Claude from a general-purpose assistant into a specialist. Rather than relying on the model’s training data alone, a skill provides explicit workflow contracts, reference material, code templates, and automated validators that load into context when the skill activates.

The large language models (LLMs) that power Claude already know the programming languages across the Salesforce ecosystem: Apex, LWC, SOQL, etc. But knowing a language and consistently producing high-quality code that adheres to best practices are two different things. A skill bridges that gap.

Each skill lives in a folder with a predictable structure:

skills/authoring-apex/
├── SKILL.md              # The execution contract
├── references/           # Decision trees, patterns, guardrails
├── assets/               # Example code
└── hooks/scripts/        # Python validators

At minimum, you need a SKILL.md, which is the entry point that Claude reads when the skill activates. Everything else is scaffolding that you add as the skill grows. Start with one SKILL.md and one reference doc, and add hooks when you want mechanical enforcement rather than relying on instructions alone.

Skills activate based on frontmatter. Frontmatter is the YAML metadata block between the ‘---’ delimiters at the top of a markdown file. It’s not rendered as content, but it’s there to be read as machine-readable configuration. In a skill, the frontmatter tells an agent when to activate: the name field identifies the skill, and the description field contains the trigger rules that the agent uses to decide whether this skill is relevant to the current task.

In the example below, the description uses a TRIGGER when / DO NOT TRIGGER when structure that gives the agent both positive and negative match criteria. This eliminates the ambiguous middle ground (like “is this SOQL query part of Apex authoring or the SOQL skill?”) that causes misfires when you only tell the model what to activate on.

---
name: authoring-apex
description: >
  TRIGGER when: user writes, edits, or reviews Salesforce Apex code —
  .cls or .trigger files including service classes, selector classes,
  trigger handlers, test classes, batch jobs, queueable classes,
  invocable methods, or Apex REST endpoints.
  DO NOT TRIGGER when: user works on LWC JavaScript, Flow XML,
  standalone SOQL queries, or running existing tests.
---

This is a simplified version. In practice, the description is more comprehensive: listing every class type (schedulable, AuraEnabled, HttpCalloutMock) and naming which skill handles each excluded case (e.g. “use authoring-lwc” for LWC JavaScript). The more specific the boundary, the fewer misfires.

When a matching task is detected, Claude loads the SKILL.md and follows the workflow. You can also trigger a skill manually with /skill-name in the prompt. Either way, the full skill context — workflow, references, templates — is injected automatically. You never paste the SKILL.md content into the conversation yourself.

The skill format itself (a SKILL.md with frontmatter, references, and assets) is an open specification maintained by Anthropic under Apache 2.0 and open to community contributions. The SKILL.md, references, and templates that you write are portable. The hooks mechanism (hooks.yaml with PostToolUse lifecycle events) is Claude Code-specific — if you use another agent runtime, you’d wire validation differently, but the skill content travels with you. This makes skills headless. Use them with Agentforce Vibes, Claude Code, Cursor, VS Code, Gemini CLI, OpenAI Codex, Windsurf, Roo Code, Goose, or any of the 30+ agents that support the open specification.

The structure of an agent skill

SKILL.md: The execution contract

This is the key, non-reorderable workflow with hard exit criteria for each phase that will be followed when the skill is used. Here is a simplified example showing the structure. Your real SKILL.md would have more detail in each phase, but the shape is the same: ordered phases with exit criteria.

Follow this workflow in order. Do not skip, merge, or reorder steps.
If blocked, stop and ask for missing context. If not applicable, mark `N/A`
with a one-line justification in the report.

## Required Inputs

Gather or infer before authoring:

- Class type (service, selector, batch, queueable, invocable, trigger handler)
- Target object(s) and business goal
- Sharing default (`with sharing` unless justified)
- Trigger framework already in use (or explicit choice)

The phases:

### Phase 1 — Author
1. **Discover project conventions** — scan for existing patterns, trigger
     framework, naming style
2. **Choose the smallest correct pattern**:

  | Need               | Pattern                              |
  | ------------------ | ------------------------------------ |
  | Business logic     | Service class                        |
  | Data access        | Selector class                       |
  | Trigger logic      | Trigger handler (framework required) |
  | Flow integration   | @InvocableMethod                     |
  | Bulk processing    | Batch Apex                           |
  | Async work         | Queueable                            |

3. **Read the matching template** from `assets/` before authoring
4. **Author with guardrails** — apply every rule in the Rules section below
5. **Generate test class** — delegate to testing skill

### Phase 2 — Validate

6. **Run code analyzer** — remediate all blocking violations; re-run until clean
7. **Execute tests** — capture pass/fail and coverage percentage

### Phase 3 — Report

8. **Report** — files, design decisions, analyzer output, test results, deploy note

If a phase doesn’t apply, Claude must document it as N/A with justification. This maps directly to Anthropic’s best practice of providing “instructions as sequential steps using numbered lists when order or completeness matters.”

references/: The knowledge library

These are your reference documents that target a specific failure mode that general training handles inconsistently. Below are some best practices from Anthropic to follow since the temptation will be to include too much information.

For an authoring-apex skill let’s explore the reference files you might consider including.

This leverages what Anthropic calls providing “context and motivation behind instructions.” The references don’t just say what to do, they explain why, enabling Claude to generalize correctly to novel situations rather than pattern-matching blindly.

As an example the anti-patterns.md could look like this. This structure is designed for practices that are likely to go wrong repeatedly. Every section follows the same BAD/GOOD formula because Claude needs to recognise and avoid independent mistakes. Learn more about these anti-patterns in the documentation.

# Apex Anti-Patterns

## Table of Contents

- [SOQL in Loops](#soql-in-loops)
- 

---

## SOQL in Loops
**Anti-pattern:** Executing SOQL queries inside a `for` or `while` loop.

**Why this fails:** Salesforce enforces a hard limit of 100 SOQL queries
per synchronous transaction. Triggers fire in batches of up to 200 records,
so a single SOQL inside a loop exhausts the limit after just 100 records.

```apex
// BAD: 1 SOQL per iteration — hits 100-query limit at record 101
for (Account acc : accounts) {
    List contacts = [SELECT Id FROM Contact WHERE AccountId = :acc.Id];
}

Fix: Query once in bulk, then iterate over results.

// GOOD: 1 SOQL total
Map<Id, List> contactsByAccount = new Map<Id, List>();
for (Contact c : [SELECT Id, AccountId FROM Contact WHERE AccountId IN :accountIds]) {
    ...
}
// add other entries to align with the anti-patterns in salesforce well-architected

For speciality examples, like transaction security policies, the markdown could follow this structure and cover one specific Apex feature that breaks all the normal rules. It needs to teach Claude when the normal rules don’t apply, what to do instead, and why. There’s no BAD/GOOD pair because the “bad” code (e.g., global without sharing) is actually the correct code in this context. Note that the example in this case refers to a worked example in the assets folder.

# Transaction Security Policy (Apex condition)

## Table of Contents

- [Overview](#overview)
- [Class shape](#class-shape)
- [Sharing](#sharing)
- [Guardrails](#guardrails)
- [Example](#example)

---

## Overview

Enhanced Transaction Security policies call Apex when Condition Builder
is not enough. Implement `TxnSecurity.EventCondition` with a single
`evaluate(SObject event)` method.

## Class shape

- Declare the class **`global`** (required for Setup selection)
- Implement `TxnSecurity.EventCondition`
- Signature: `global Boolean evaluate(SObject event)`

## Sharing

These classes are **not** user-facing controllers. Use `without sharing`.
The normal `with sharing` default does NOT apply here -- applying it
will break the policy silently.

## Guardrails

- One bulk SOQL query max (no SOQL in loops -- governor limits still apply)
- Return `true` to block the transaction, `false` to allow

## Example

See assets/transaction-security-policy.cls

assets/: Few-shot by example

Anthropic highlights that one of the most reliable ways to steer Claude’s output format, tone, and structure is with implicit few-shot examples. Few-shot just means that you provide a few good examples as a template. Each template embeds the non-negotiable requirements: ApexDoc comments, bulk-safe logic, explicit sharing declarations, CRUD/FLS enforcement, and dependency injection for testability. When Claude adapts a template, it inherits these properties by construction rather than needing to recall them from training.

For Apex specifically, the world is your oyster here, but let’s explore some good use cases.

Each template isn’t just “how to write X.” It’s a pre-loaded, few-shot example with the non-negotiable requirements baked into the structure.

hooks.yaml and validator scripts: The safety net

A hook is a shell command that Claude Code runs automatically at a specific lifecycle moment. You configure them in hooks.yaml. The script is whatever that command executes, typically a Python validator that inspects the tool’s input/output and returns structured feedback.

Hooks close the feedback loop without human intervention. Claude writes something, the hook evaluates it immediately, and Claude sees the result in the same conversation turn. The correction happens in context while Claude still has the full problem loaded.

The hooks file wires the validator to the right moment. For example, a preflight check runs when the user submits a prompt.

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "python3 ${CLAUDE_SKILL_DIR}/hooks/scripts/preflight-apex-check.py",
            "timeout": 10000
          }
        ]
      }
    ]
  }
}

Or a PostToolUse hook fires after every file write or edit.

---
name: authoring-apex
description: ...
hooks:
  PostToolUse:
    - matcher: "Write|Edit"
      hooks:
        - type: command
          command: "python3 ${CLAUDE_SKILL_DIR}/hooks/scripts/preflight-apex-check.py"
          timeout: 90000
---

Claude Code delivers the hook context as JSON on stdin. The payload shape varies by lifecycle event: PostToolUse includes the tool name, input parameters, and output; UserPromptSubmit includes the user’s prompt text.

Consider a preflight check for your development team. The SKILL.md workflow says “run Code Analyzer” in the validate phase. But what if Code Analyzer isn’t installed or is the wrong version? Without a hook, Claude attempts the command mid-workflow, gets an error, and the developer has to diagnose a missing plugin. With a hook, the check runs the moment Apex work begins.

import json, subprocess, sys

MINIMUM_VERSION = "5.5.0"

def parse_version(v):
    """Parse version string into comparable tuple."""
    return tuple(int(x) for x in v.split(".")[:3])

def main():
    hook_input = json.load(sys.stdin)
    prompt = hook_input.get("prompt", "").lower()

    # Only check when Apex work is likely
    apex_keywords = ["apex", ".cls", "class", "trigger", "service", "selector", "batch"]
    if not any(kw in prompt for kw in apex_keywords):
        sys.exit(0)

    try:
        result = subprocess.run(
            ["sf", "plugins", "--json"],
            capture_output=True, text=True, timeout=15
        )
        if result.returncode == 0:
            plugins = json.loads(result.stdout)
            for plugin in plugins:
                if plugin.get("name") == "@salesforce/plugin-code-analyzer":
                    version = plugin.get("version", "0.0.0")
                    if parse_version(version) >= parse_version(MINIMUM_VERSION):
                        sys.exit(0)
                    else:
                        print("=== Preflight Check ===")
                        print(f"  [WARNING] Code Analyzer {version} is below minimum {MINIMUM_VERSION}.")
                        print(f"  Install: sf plugins install @salesforce/plugin-code-analyzer@latest")
                        sys.exit(0)
    except (FileNotFoundError, subprocess.TimeoutExpired, json.JSONDecodeError):
        pass

    print("=== Preflight Check ===")
    print("  [WARNING] Salesforce Code Analyzer is not installed.")
    print("  The authoring-apex skill requires it for validation.")
    print("  Install: sf plugins install @salesforce/plugin-code-analyzer")
    sys.exit(0)

if __name__ == "__main__":
    main()

The developer is told up front that their toolchain is incomplete rather than discovering it mid-workflow when the validate phase fails. 

That’s what separates an instruction from a hook. The SKILL.md says “run Code Analyzer in the validate phase.” The hook proves that it can run before the workflow even starts. Instructions are aspirational; hooks are mechanical. You want both capabilities.

Why agent skills need human and mechanical validation

Skills are not a silver bullet. They dramatically narrow the probability distribution of Claude’s outputs, but they operate on a fundamentally non-deterministic system. Two things work against you:

You can’t fully specify behavior. No matter how precise your SKILL.md, real-world requirements contain novel combinations that aren’t covered by templates or references. Claude must still make judgement calls and newer models interpret prompts more literally, so slight underspecification produces inconsistent results. If your SKILL.md says “add error handling” without specifying the pattern, you’ll get different approaches each time.

Outputs are probabilistic. Even with structured prompts, role assignment, examples, and validation hooks, you’re optimizing a hit rate. You move from 60% correct to 95% correct — a massive improvement —  but the remaining 5% is why validators exist. This is why the architecture includes both preventive measures (structured prompts, references, templates) and detective measures (automated validators, scoring rubrics, blocking on errors). The skill assumes Claude will sometimes get it wrong and builds correction into the workflow rather than pretending perfection is achievable.

Getting started

Anthropic’s golden rule for prompts applies directly: “Show your prompt to a colleague with minimal context and ask them to follow it. If they’d be confused, Claude would be too.” A well-built agent skill is unambiguous to both a human reader and the model.

You don’t need to build a skill from scratch. Anthropic provides a skill-creator skill that walks you through the full process: capturing intent, writing the SKILL.md, creating test cases, running evals, and iterating until the output meets your bar. Install it and tell Claude what you want the skill to do; it handles the scaffolding, interviews you on edge cases, and generates a working draft you can refine.

If you’d prefer to work from an existing Salesforce-specific example, the Agentforce Vibes skill library includes production-ready skills for Apex, LWC, Flow, and more. Install them, use them, and look at how they’re structured; they follow similar patterns described in this post. The Agentforce Vibes skill library puts rules and rationale inline in the SKILL.md and uses complete Apex source files as templates rather than Markdown reference docs. This keeps the skill self-contained in fewer files, though it trades modularity: updating one rule means editing the main workflow file. 

The skill grows with your needs. Start with the skill-creator or an existing skill, customize what doesn’t fit your project, and build new skills only when you have a gap not covered by existing ones.

Resources

• GitHub: Salesforce’s collection of agent skills
• Trailhead: Prompt Fundamentals
• Trailhead: Prompt Engineering Techniques
• GitHub: Anthropics Skill Creator
• Documentation: Claude Code 

About the author

Dave Norris is a Developer Advocate at Salesforce. He’s passionate about making technical subjects broadly accessible to a diverse audience. Dave has been with Salesforce for over a decade, has over 40 Salesforce and MuleSoft certifications, and became a Salesforce Certified Technical Architect in 2013.

The post Build a Salesforce Agent Skill with Claude Code appeared first on Salesforce Developers Blog.

In this post, we’ll show you how to integrate Discovery Framework, an Omniscript-based digital form engine in Agentforce for Financial Services, directly into an Agentforce agent. 

By the end, you’ll have a working solution that renders structured forms inside the Agent window, replacing error-prone conversational capture with a reliable, UI-driven experience. This solution is for Salesforce Developers and architects building Agentforce experiences for the financial services industry. It requires familiarity with Apex, Lightning Web Components (LWC), and Omniscript basics.

The problem: Conversational capture falls short

In financial services, a typical transaction dispute intake process has five stages:

1. Identifying the impacted customer and account
2. Identifying the impacted transaction with enrichment for ease of identification
3. Capturing the reason for dispute and disputed amount
4. Creating a case for dispute intake
5. Capturing a dispute questionnaire

For high-volume contact center agents, Stage 5 is the hardest. Questionnaires are dynamic, changing based on payment network rules and transaction type.

A pure conversational experience creates real problems:

• Users can’t navigate backwards to update responses
• Special UI patterns (radio buttons, checkboxes) don’t translate to text
• Any AI paraphrase of a question risks capturing the wrong answer
• Questions must appear in strict order, with required/optional labels intact
• Longer questionnaires need save-and-resume support

The screen recording below shows an agent in conversational mode, asking dispute questionnaire questions one at a time.

The solution: Discovery Framework in agents

Agentforce Financial Services provides Discovery Framework, a feature that creates digital forms to collect and validate data while avoiding time-consuming, error-prone manual methods.

This framework was originally built for non-agentic, UI-driven scenarios. When you bring it into agents, you give users navigable, structured forms that address all five problems discussed above.

Key benefits:

• Navigation between questions with ability to update responses
• Proper rendering of radio buttons, checkboxes, and other form controls
• Exact question display without AI interpretation
• Correct ordering and required/optional field marking
• State persistence for incomplete forms

The framework renders a structured form directly in the agent output, with no conversational questions and no AI paraphrase risk. The screen recording below shows an agent in a Discovery Framework experience with structured form fields, radio buttons, and navigation controls embedded in the Agent interface.

Prerequisites: This experience is available only in UI-based channels — Lightning Experience, Messaging for Experience Sites, and third-party sites. It is not supported in Voice or Short Message Service (SMS) channels, where complex UI elements can’t be displayed.

Technical architecture overview

The integration uses Agentforce’s support for custom Lightning types and custom LWC components. Here’s the component architecture:

1. Discovery Framework Omniscript: The digital form definition with assessment questions
2. Custom LWC component: Wraps the Omniscript and handles data binding
3. Apex class: Executes business logic when the Agent action runs
4. Custom Lightning type: Associates the Apex output type with the LWC renderer
5. Agent action: Uses the Apex class as its reference action and the custom Lightning type as its output renderer

When an Agent action executes via a subagent, the custom Lightning type renders in the output, displaying the Discovery Framework Omniscript directly in the Agent window.

The architecture diagram below illustrates the runtime flow: a subagent triggers an agent action, which then executes Apex class business logic and renders the custom Lightning type LWC component, wrapping the Discovery Framework Omniscript in the Agent interface.

To learn more about the underlying capability, see Enhance the Agent UI with Custom LWCs and Lightning Types.

Steps to implement Discovery Framework in Agentforce

The implementation involves five steps. You start by creating a Discovery Framework Omniscript, then build the Apex classes that define the agent action’s business logic. Next, you create a Lightning web component to wrap and render the Omniscript inside the Agent window. You then configure a custom Lightning type to associate the Apex output with the LWC, and finally, you set up the agent action, which uses the Apex class as the reference action and the Lightning type as the output render. Let’s take a detailed look.

Step 1: Create a Discovery Framework Omniscript

First, we’ll need to create a series of assessment questions, then add them as steps in an Omniscript (a drag-and-drop digital form builder in Discovery Framework). See the Discovery Framework documentation for a step-by-step walkthrough.

For this example, we’ll create two sample dispute questions of type Radio Button with Yes/No responses.

Omniscript properties:

• Type: FinancialServices
• SubType: DisputeQuestionnaire
• Language: English

Important: Configure the Omniscript for a smaller viewport, so it renders correctly inside the agent window. Open the Omniscript properties panel and set the Viewport width to a maximum of 600px. Enable Responsive Mode if available. This ensures that the form fits within the agent output area without scrolling issues.

Navigation configuration:

• Enable Allow Save for Later to let users save incomplete forms
• Enable Show Progress Indicator to display form completion status
• Configure step-by-step navigation with Previous and Next buttons
• Select Hide step chart in Step Chart Options

For detailed instructions on creating Discovery Framework Omniscripts, see the Discovery Framework documentation.

To learn more, watch the video: Discovery Framework for Financial Services Cloud 

Step 2: Create Apex classes

Next, we’ll create an Apex class that captures the business logic for the agent action. In this example, the class creates a real Case record and returns the case number and ID. You’ll want to adapt the logic to your business needs.

The following Apex classes define the input, output, and business logic for the agent action. Each class plays a specific role in handling the incoming dispute request, structuring the output data, and executing the case creation logic.

Here is sample code for the Apex classes noted above.

DisputeQuestionnaireInput.cls
This class serves as the input of the agent action and contains a disputeRequest String field.
See sample code.

DisputeQuestionnaireOutput.cls
This class serves as the output of the agent action and contains a DisputeQuestionnaireCaseOutput object.
See sample code.

DisputeQuestionnaireCaseOutput.cls
This class contains the caseNumber and caseId fields as Strings.
See sample code.

DisputeQuestionnaireAction.cls
This class has the business logic that is executed by the agent to create a case and add it to a DisputeQuestionnaireCaseOutput object which is embedded inside a DisputeQuestionnaireOutput object.
See sample code.

Step 3: Create a Lightning web component

Next, we’ll create a Lightning Web Component that wraps the Discovery Framework Omniscript from Step 1. 

The displayDisputeQuestionnaire Lightning web component wraps the Discovery Framework Omniscript and renders it inside the Agent window. This component includes an HTML, a JavaScript and an XML file. The HTML defines the component structure, the JavaScript handles data binding and event subscription, and the metadata XML registers the component as an Agentforce output target.

Here is a sample code for the displayDisputeQuestionnaire Lightning web component.

displayDisputeQuestionnaire.html
This HTML component wraps the Discovery Framework Omniscript created in Step 1, using <omnistudio-omnistudio-standard-runtime-wrapper>.

<template>
   <div class="slds-card">
           <div class="slds-card">
               <div class="slds-modal__content">
                   <omnistudio-omnistudio-standard-runtime-wrapper
                           type="FinancialServices"
                           subtype="DisputeQuestionnaire"
                           language="English"
                           prefill={prefill}>
                  </omnistudio-omnistudio-standard-runtime-wrapper>
               </div>
           </div>
   </div>
</template>

Note: We’ll use the same Type, SubType, and Language values in the HTML as configured in Step 1, as shown in above code snippet.

See sample code.

displayDisputeQuestionnaire.js
This is the JavaScript for the LWC, supporting the above HTML.

import { LightningElement, api, track } from 'lwc';
import { NavigationMixin } from 'lightning/navigation';
import pubsub from 'omnistudio/pubsub';

export default class DisplayDisputeQuestionnaire extends NavigationMixin(LightningElement) {
   _value;
   @api
   get value() {
       return this._value;
   }
   set value(value) {
       this._value = value;
   }
   @track prefill; assessmentId; caseNumber; caseId;
   connectedCallback() {
       this.caseNumber = this.value?.caseOutput?.caseNumber || null;
       this.caseId = this.value?.caseOutput?.caseId || null;
       pubsub.register('omniscript_action', {
           data: this.handleOmniAction.bind(this),
       });
   }
   handleOmniAction(data) {
       this.assessmentId = data?.assessmentId || null;
   }
}

See sample code.

displayDisputeQuestionnaire.js-meta.xml
This is the js-meta.xml for the LWC, which captures the target as AgentforceOutput.
See sample code.    

Key implementation details:

• The component subscribes to omniscript_action events to capture form completion
• The prefill property allows pre-populating form fields with contextual data
• The value property receives output from the Apex class
• Navigation handles redirecting users to the created Case record after completion
• The lightning__AgentforceOutput target in the metadata XML is required for the component to render in Agentforce

Step 4: Create a custom Lightning type

Next, we’ll create a custom Lightning type to associate the Apex output types from Step 2 with the LWC from Step 3.

The displayQuestionnaireResponse Lightning type defines the custom Lightning type. This Lightning type includes the JSON files, which tells Agentforce how to map your Apex output to the LWC renderer and validates the data structure at runtime.

Here is a sample code for the displayQuestionnaireResponse Lightning type.

renderer.json
Therenderer.json file maps the root output ($) to your LWC component.
See sample code.

schema.json
The schema.json file defines the data type structure matching your Apex output class.
See sample code.

For detailed setup instructions, see Enhance the Agent UI with Custom LWCs and Lightning Types.

Step 5: Create an agent action

Finally, we’ll create an agent action that uses the Apex classes from Step 2 for execution and the custom Lightning type from Step 4 for output rendering.

Agent action configuration steps:

1. Navigate to Setup > Agentforce > Agent Actions
2. Click New Agent Action
3. Set Reference Action Type to Apex
4. Select DisputeQuestionnaireAction as the Apex class
5. Set Output Rendering to the DisputeQuestionnaireResponse custom Lightning type
6. Configure input parameters to match the Agent Topic context
7. Add instructions for when the agent should invoke this action

Example subagent instruction:

When a customer wants to dispute a transaction, execute the Process Dispute Request action to capture detailed dispute information through a structured questionnaire. When this agent action fires as part of a subagent, the Discovery Framework Omniscript renders directly in the Agent window.

The screenshot below shows a sample agent action configuration in Agent Builder, with the disputeQuestionnaireResponse Lightning type created in Step 4 set as the output renderer.

Rendering the Discovery Framework Omniscript in an Agent Script action

If your Agent is built using Agent Script, you can also render the Discovery Framework Omniscript in an action which is part of the Agent Script. 

For the dispute example, the code snippet below would help to render the output of the action using the Discovery Framework Omniscript.

outputs:
    caseOutput: object
        label: "Case Output"
        description: "The dispute case details including case number and case Id"
        complex_data_type_name: "c__DisputeQuestionnaireResponse"
        developer_name: "caseOutput"
        is_displayable: True
        filter_from_agent: False

Testing the integration

The following are steps to test the complete integration.

1. Create a subagent that invokes your new agent action
2. Start a conversation with the agent
3. Trigger the dispute intake flow
4. Verify that the Discovery Framework form renders in the Agent interface
5. Complete the questionnaire and verify that data is captured correctly
6. Confirm that the Case record is created with the captured information

Troubleshooting tips:

• If the form doesn’t render, verify that the Omniscript Type, SubType, and Language match exactly in both the Omniscript definition and LWC component
• If data doesn’t pass correctly, check that @AuraEnabled annotations are present on all Apex output class properties
• If navigation fails, ensure that the LWC component’s js-meta.xml includes the lightning__AgentforceOutput target

Extending this pattern

This mechanism isn’t limited to Discovery Framework Omniscripts; you can apply it to any Omniscript in your org, as long as the user experience fits within a compact viewport.

Additional use cases:

• Loan application intake: Multi-step loan applications with document upload
• Insurance claims processing: Complex claim forms with conditional logic
• Customer onboarding: Know Your Customer (KYC) questionnaires with validation rules
• Financial planning: Goal-setting forms with calculation logic

The key requirement: design Omniscripts for responsive display within the agent interface viewport constraints.

Conclusion

With these five steps, you can replace fragile conversational question-and-answer flows with a reliable, structured form experience — directly inside Agentforce. This approach combines the conversational intelligence of Agentforce with the rich data collection capabilities of Discovery Framework, delivering the best of both worlds.

By following this approach, you maintain accurate data capture without AI hallucination, while giving users intuitive, navigable forms that feel natural within the Agent conversation flow. Try it out and share what you build. Post questions on the Salesforce Developer Community or Stack Overflow.

Resources

• Documentation: Transaction Dispute Management Intake in Financial Services Cloud
• Documentation: Discovery Framework
• Video: Discovery Framework for Financial Services Cloud
• Developer Guide: Enhance the Agent UI with Custom LWCs and Lightning Types 

About the author

Aditya Bhansali is a Lead Member of Technical Staff in the Agentforce Financial Services organization, where he leads Salesforce engineering teams in building enterprise products on the Salesforce Platform. He shares updates on LinkedIn.

The post Capture Rich Data within Agent Conversations by Using Discovery Framework appeared first on Salesforce Developers Blog.

In this post, we’ll explore the benefits of working with MCP servers, then we’ll focus specifically on how to configure Claude for MCP, and finally, we’ll share some sample use cases.

What are Salesforce Hosted MCP Servers?

The Model Context Protocol (MCP) is an open standard that lets agentic tools like Postman or AI agents like Claude interact with external tools and data sources from various vendors. In the context of Salesforce, it means that you can run SOQL queries, modify records, and execute actions directly from a third-party agent. In other words, you don’t have to log into Salesforce and you’re no longer limited to using Agentforce to interact with your data from agents.

In order to perform operations on Salesforce with agents, you use MCP servers that expose tools and prompts. You can configure and host your own Salesforce MCP Server or use the Salesforce Hosted MCP Servers that are now generally available. We’ll only focus on the latter in this post since this approach saves you from the burden of hosting, securing, and maintaining the server.

Standard Salesforce Hosted MCP Servers

In addition to the Salesforce DX MCP Server used for development purposes, Salesforce offers a number of standard hosted MCP servers out of the box for daily operations. These allow agents to access our Headless 360 portfolio ranging from the Salesforce Platform, Data 360, Tableau, and MuleSoft. You can also connect directly to the Slack hosted MCP server.

Custom Salesforce Hosted MCP Servers

If the standard hosted MCP servers do not fit your needs, you can also create custom hosted MCP servers. This option gives you granular control over the tools and prompts that the server exposes, and lets you add custom tools and prompts.

You can build custom MCP tools from the following types:

• Apex Action: Expose invocable Apex Actions (@InvocableMethod annotated Apex methods) as MCP tools
• Lightning Flows: Expose autolaunched flows as MCP tools
• Apex REST: Expose custom Apex REST endpoints as MCP tools.
• AuraEnabled: Expose @AuraEnabled (see docs) annotated Apex methods as MCP tools
• Named Query API: Expose Named Query APIs as MCP tools to provide access to specific parametrized SOQL queries

You can also create custom MCP prompts from Flex prompt templates built with Prompt Builder. Using MCP prompts in an agent lets you call those parametrized prompt templates as commands. For example, if you expose a “Generate Personalized Schedule” prompt template, you can run the following command in Claude (you don’t have to type the full tool name, autocomplete provides the prompt description).

Now that we’ve seen the benefits of working with the Salesforce Hosted MCP Servers, let’s look at how to configure Claude to leverage them.

Configure Claude to work with Salesforce Hosted MCP Servers

Configuring Claude to connect with the hosted MCP servers is a three-step process:

1. Activate Salesforce Hosted MCP Servers
2. Create an external client app
3. Connect Claude to the Salesforce Hosted MCP Servers

Note: Claude is available in different surfaces. In this post, we’ll cover Claude Code (the CLI experience) and Claude Desktop (the desktop app).

As part of the setup, you’ll need to set up an external client app (ECA) so that Claude can authenticate via OAuth 2.0 with Salesforce to use the hosted MCP servers. An ECA tells Salesforce which external applications are allowed to connect to your org and what they’re allowed to do. Standard Salesforce user access permissions apply in addition to the app-level access checks.

Step 1: Activate Salesforce Hosted MCP Servers

The standard Salesforce Hosted MCP Servers are inactive by default. Follow these steps to activate them:

1. In Setup, search for MCP Servers.
2. Click the Salesforce Servers tab.
3. Repeat the following steps for each server that you want to activate:
   1. Click on the server name.
   2. Click Activate.
   3. Copy the server’s API Name without the platform. suffix and the Server URL. We’ll need those in the next step.

Step 2: Create an external client app

Follow these steps to create an ECA:

1. In Setup, search for External Client App Manager and click New External Client App.
2. Under Basic Information, fill in the required fields:
   

   Field name	Field Value
   App Name	Claude MCP Client
   API Name	Claude_MCP_Client (auto-filled)
   Contact Email	Your email address

3. Expand the API (Enable OAuth Settings) section, check Enable OAuth and configure the following app settings:
   1. Callback URL with either:
      • https://claude.ai/api/mcp/auth_callback for Claude Desktop
      • http://localhost:38000/callback for Claude Code. Use this if you only have an Anthropic API key provisioned for you by your enterprise. Otherwise, we recommend setting up connectors in your Claude.ai account, then using them in Claude Code.
   2. Selected OAuth scopes:
      • Perform requests at any time (refresh_token, offline_access)
      • Access Salesforce Hosted MCP Servers (mcp_api)
4. Under Security, apply the following:
   1. Uncheck these boxes:
      • Require secret for Web Server Flow
      • Require secret for Refresh Token Flow
   2. Check these boxes:
      • Require Proof Key for Code Exchange (PKCE) extension for Supported Authorization Flows
      • Issue JSON Web Token (JWT)-based access tokens for named users 

        At this point, your ECA configuration should look like this:
        

5. Click Create to save the app.

Now that the app is created, you need to retrieve the connection information:

1. Navigate to the Settings tab and expand OAuth Settings.
2. Click Consumer Key and Secret.
3. Complete the authentication dialog with the verification code.
4. Copy the Consumer Key and Consumer Secret.

Step 3: Connect Claude to the Salesforce Hosted MCP Servers

The last step of the integration process is to configure the Salesforce MCP Server connections in Claude.

The instructions differ depending on whether you use Claude Code or Desktop (in particular the ECA callback URL). We’ll dive into how to configure Claude Code here, and you can find instructions for Claude Desktop in the docs.

Instructions for Claude Code

1. Run the following command in a terminal where:
   – MY_MCP_SERVER_NAME is replaced by the name of the server that you’re connecting to. You’re free to enter any value, but we recommend that you stick to the following convention: salesforce- followed by the API name of the MCP server. For example: salesforce-sobject-all for the sobject-all server.
   –MY_MCP_SERVER_URL is replaced by the MCP server URL that you copied earlier. Pay attention to the fact that server URLs differ depending on whether you are connecting to a production or sandbox/scratch org.
   –MY_ECA_CONSUMER_KEY is replaced by the ECA consumer key you’ve obtained in the previous step.

   claude mcp add --transport http MY_MCP_SERVER_NAME MY_MCP_SERVER_URL --callback-port 38000 --client-id "MY_ECA_CONSUMER_KEY" --client-secret
   

   For example, to install the sobject-all MCP server from a scratch org, run:

   claude mcp add --transport http salesforce-sobject-all https://api.salesforce.com/platform/mcp/v1/sandbox/platform/sobject-all --callback-port 38000 --client-id "XXX" --client-secret
   

2. When prompted to do so, paste the ECA consumer secret you’ve obtained previously.
3. Launch Claude Code with the claude terminal command.
4. Run /mcpto list the MCP servers.
5. Select your newly added MCP server from the list (it is annotated with a “needs authentication” message).
6. Select Authenticate.
7. When prompted, allow access to the MCP server.

Once you authorize the connection, Claude Code lists your MCP server as “connected.”

Some practical examples of Claude working with Salesforce Hosted MCP Servers

Let’s assume that you work at Coral Cloud Resorts, a fictitious resort that offers tailored experience to their guests using Headless 360. You can use Claude and our MCP servers for a number of tasks.

Analyzing daily bookings

Let’s imagine that as the resort manager, you want to review customer bookings on a weekly basis.

In traditional Salesforce operations, you would build a report and then generate a dashboard out of it. To access it, you would need tens of clicks in Lightning Experience.

With Claude, all you have to do is use natural language, such as “Show me a report of daily bookings. Use the start date to group bookings,” to get the trend of your bookings with some additional insights that would require that you set up Tableau reports. You can also drill down by interacting further with Claude with follow up questions.

❯ Show me a report of daily bookings. Use the start date to group bookings.                                                                                      

⏺ I'll query the bookings grouped by start date now.
  
  Called salesforce-sobject-all (ctrl+o to expand)

⏺ Let me find the correct booking object name in this org.

  Called salesforce-sobject-all 2 times (ctrl+o to expand)

⏺ The date field is Date__c. Let me query the bookings grouped by that field.

  Called salesforce-sobject-all (ctrl+o to expand)

⏺ Here's the daily bookings report:

  ┌────────────┬──────────┬──────────────┬───────────────┐
  │    Date    │ Bookings │ Total Guests │ Total Revenue │
  ├────────────┼──────────┼──────────────┼───────────────┤
  │ 2026-05-10 │ 347      │ 1,096        │ $144,918      │
  ├────────────┼──────────┼──────────────┼───────────────┤
  │ 2026-05-11 │ 316      │ 1,026        │ $132,874      │
  ├────────────┼──────────┼──────────────┼───────────────┤
  │ 2026-05-12 │ 325      │ 996          │ $131,169      │
  ├────────────┼──────────┼──────────────┼───────────────┤
  │ 2026-05-13 │ 347      │ 1,050        │ $135,670      │
  ├────────────┼──────────┼──────────────┼───────────────┤
  │ 2026-05-14 │ 285      │ 887          │ $106,410      │
  ├────────────┼──────────┼──────────────┼───────────────┤
  │ Total      │ 1,620    │ 5,055        │ $651,041      │
  └────────────┴──────────┴──────────────┴───────────────┘

  A 5-day window with fairly consistent volume (~315–347/day), except May 14 which is notably lower at 285 bookings and $106K revenue — about 18% below the average of the other days.

Reading data is nice, but agents can take things further with MCP and take action.

Mass cancelling bookings using a custom server and a flow

Running a resort is no simple feat, and there can be unforeseen events (such as jellyfish invasions) that lead to booking cancellations. Should such unlikely scenarios occur, the resort staff is left scrambling to mass cancel bookings with hundreds of clicks, which is a repetitive and error-prone process.

Instead, you can leverage Headless 360 and expose our existing Cancel Booking flow as a custom hosted MCP server.

Note: in this basic example, you use a flow to cancel a single booking at a time. However, in production, we recommend creating a dedicated flow to process cancellations in bulk to reduce API calls and scale.

Once the custom MCP server is added to Claude, you can run the cancellations with a prompt that combines multiple calls to our various MCP servers. Note how the prompt is structured to keep the human in the loop: you validate the data before running the cancellation batch.

❯ Identify all bookings for tomorrow morning for the "Surf's Up Lesson" and "Tropical Snorkel Adventure" experience.            
  Let me verify the list then cancel them with the following cancellation reason: "Dear guest, we're sorry to cancel your aquatic experience due to a jellyfish invasion. You'll be refunded shortly."                                                  

⏺ Let me query tomorrow's morning bookings for those two experiences.

  Called salesforce-sobject-all (ctrl+o to expand)

⏺ Here are the 12 bookings found for tomorrow morning (May 13) — please verify before I proceed with cancellations:

  Surf's Up Lesson (9:00 AM – 10:00 AM) — 5 bookings

  ┌─────┬────────────┬───────────────────┬──────────────────────────────┬────────┐
  │  #  │  Booking   │       Guest       │            Email             │ Guests │
  ├─────┼────────────┼───────────────────┼──────────────────────────────┼────────┤
  │ 1   │ B-00001602 │ Ferne Ackwood     │ ferneackwood@example.com     │ 1      │
  ├─────┼────────────┼───────────────────┼──────────────────────────────┼────────┤
  │ 2   │ B-00001603 │ Deeyn Dumsday     │ deeyndumsday@example.com     │ 2      │
  ├─────┼────────────┼───────────────────┼──────────────────────────────┼────────┤
  │ 3   │ B-00001604 │ Alessandro Krates │ alessandrokrates@example.com │ 2      │
  ├─────┼────────────┼───────────────────┼──────────────────────────────┼────────┤
  │ 4   │ B-00001605 │ Quinlan Soppit    │ quinlansoppit@example.com    │ 2      │
  ├─────┼────────────┼───────────────────┼──────────────────────────────┼────────┤
  │ 5   │ B-00001606 │ Osborn Libbe      │ osbornlibbe@example.com      │ 3      │
  └─────┴────────────┴───────────────────┴──────────────────────────────┴────────┘

  Tropical Snorkel Adventure (9:00 AM – 1:00 PM) — 7 bookings

  ┌─────┬────────────┬────────────────────┬───────────────────────────────┬────────┐
  │  #  │  Booking   │       Guest        │             Email             │ Guests │
  ├─────┼────────────┼────────────────────┼───────────────────────────────┼────────┤
  │ 6   │ B-00000972 │ Gay Ragge          │ gayragge@example.com          │ 1      │
  ├─────┼────────────┼────────────────────┼───────────────────────────────┼────────┤
  │ 7   │ B-00000973 │ Baxie Donaher      │ baxiedonaher@example.com      │ 5      │
  ├─────┼────────────┼────────────────────┼───────────────────────────────┼────────┤
  │ 8   │ B-00000974 │ Munmro Semken      │ munmrosemken@example.com      │ 3      │
  ├─────┼────────────┼────────────────────┼───────────────────────────────┼────────┤
  │ 9   │ B-00000975 │ Ira Kemble         │ irakemble@example.com         │ 1      │
  ├─────┼────────────┼────────────────────┼───────────────────────────────┼────────┤
  │ 10  │ B-00000976 │ Suzie Patience     │ suziepatience@example.com     │ 3      │
  ├─────┼────────────┼────────────────────┼───────────────────────────────┼────────┤
  │ 11  │ B-00000977 │ Stephine Suttill   │ stephinesuttill@example.com   │ 1      │
  ├─────┼────────────┼────────────────────┼───────────────────────────────┼────────┤
  │ 12  │ B-00000978 │ Christyna Gianelli │ christynagianelli@example.com │ 1      │
  └─────┴────────────┴────────────────────┴───────────────────────────────┴────────┘

  Total: 12 bookings, 25 guests affected.

  All are currently Confirmed. Confirm and I'll cancel all 12 with the jellyfish invasion message.

✻ Cogitated for 18s

❯ confirmed                                                                                                                      

⏺ Let me load the Cancel Booking flow tool and cancel all 12 bookings in parallel.

⏺ Cancelling all 12 bookings in parallel now.

  Called salesforce-coral-cloud-flows 12 times (ctrl+o to expand)

⏺ All 12 bookings cancelled successfully. Here's the summary:

  Surf's Up Lesson — 5 cancelled

  ┌────────────┬───────────────────┐
  │  Booking   │       Guest       │
  ├────────────┼───────────────────┤
  │ B-00001602 │ Ferne Ackwood     │
  ├────────────┼───────────────────┤
  │ B-00001603 │ Deeyn Dumsday     │
  ├────────────┼───────────────────┤
  │ B-00001604 │ Alessandro Krates │
  ├────────────┼───────────────────┤
  │ B-00001605 │ Quinlan Soppit    │
  ├────────────┼───────────────────┤
  │ B-00001606 │ Osborn Libbe      │
  └────────────┴───────────────────┘

  Tropical Snorkel Adventure — 7 cancelled

  ┌────────────┬────────────────────┐
  │  Booking   │       Guest        │
  ├────────────┼────────────────────┤
  │ B-00000972 │ Gay Ragge          │
  ├────────────┼────────────────────┤
  │ B-00000973 │ Baxie Donaher      │
  ├────────────┼────────────────────┤
  │ B-00000974 │ Munmro Semken      │
  ├────────────┼────────────────────┤
  │ B-00000975 │ Ira Kemble         │
  ├────────────┼────────────────────┤
  │ B-00000976 │ Suzie Patience     │
  ├────────────┼────────────────────┤
  │ B-00000977 │ Stephine Suttill   │
  ├────────────┼────────────────────┤
  │ B-00000978 │ Christyna Gianelli │
  └────────────┴────────────────────┘

  All 25 affected guests have been notified and will be refunded.

These are two basic examples of what you can do with Claude and Salesforce Hosted MCP Servers, ranging from standard actions to custom scenarios.

Conclusion

You’ve now had a glimpse at how Salesforce Hosted MCP Servers let you use agents like Claude to interact with Headless 360 without requiring users to log into Lightning Experience. You’ve learned how to configure Claude Code and Claude Desktop to connect with Salesforce Hosted MCP Servers, and you saw some practical examples of what can be done with those servers.

Now, it’s your turn to put Headless 360 to the test by building custom tools and configuring MCP servers to power your business.

Resources

• Documentation: Salesforce Hosted MCP Servers
• Documentation: Claude Code
• Postman collection: Salesforce Hosted MCP Servers

About the Author

Philippe Ozil is a Principal Developer Advocate at Salesforce, where he focuses on the Salesforce Platform. He writes technical content and speaks frequently at conferences. He is a full-stack developer and enjoys working with APIs, DevOps, robotics, and VR projects. Follow him on X, LinkedIn, and Bluesky, and check out his GitHub projects.

The post Connect Claude with Salesforce Hosted MCP Servers appeared first on Salesforce Developers Blog.

As the Salesforce Developer Advocacy team lead, I recently posted on the future of the Salesforce Developer role. In that post, I argued that AI is reshaping our work around system design, quality engineering, and cross-system thinking. Headless 360 is the platform shift that turns those arguments into daily practice.

The platform provides context and capability

The LLM provides intelligence. The platform provides everything else: identity, access, capabilities, governance, and the context that makes intelligence useful. When an agent calls an MCP tool, it’s invoking accumulated state. Without it, the agent has nothing to reason over, no permissions to inherit, no business process to follow.

This was always true. Every customer record, every workflow, every approval chain, every permission boundary your org has accumulated – that’s not overhead. That’s the business itself, encoded in a system that enforces it.

That combination of intelligence from the model and everything else from the platform is what separates a prototype from a production system. It’s also what makes Headless 360 different from “just expose some APIs”.

Two kinds of MCP tools

When we talk about MCP tools in the Salesforce ecosystem, we mean two things:

• MCP tools for coding agents. These enable developers to connect Claude Code, Cursor, or any MCP-compatible coding agent to their org and read metadata, generate Apex, run tests, or deploy through the pipeline. The consumer is a developer building software.
• MCP tools for business agents. These enable users to access Salesforce capabilities from any MCP-compatible client: including Slack, ChatGPT, and Claude, among others. These MCP tools also allow agents to access Salesforce capabilities autonomously. The consumer is an agent serving an end user.

Both kinds of MCP tools share the same open surface and the same trust layer. The difference is who’s calling and why. Headless 360 covers both, and both have implications for the expertise you’ve built.

Headless 360 changes how the platform can be consumed. Coding agents can now reach into your org through MCP tools to build and deploy. Business agents can call the same platform to serve customers in Slack, ChatGPT, Claude, or any MCP-compatible client. The Salesforce browser-based UI is one surface among many..

How software is built, delivered, and used is changing all at once

Agentic AI is changing three things simultaneously.

How software is written. Agents are becoming the development team. The first draft of an Apex class, an LWC, a test suite, or an Agentforce agent arrives in seconds. Your job shifts from writing code to specifying intent, evaluating output, and deciding what fits your existing system design.

How quickly it is delivered. Implementations compress from months to days. When an agent can execute against a development org, with human review gates before anything reaches production, the bottleneck moves from execution to clarity of intent.

How it is used. Humans will use Salesforce from anywhere, not necessarily the Salesforce UI. A sales rep gets a customer summary in Slack. A support agent resolves a case in Teams. And increasingly, the consumer isn’t a human at all; it’s a business agent reasoning over your data. The destination isn’t a Lightning Experience tab. It’s wherever the work is already happening.

These three changes are happening simultaneously, and they put pressure on the platform itself. Headless 360 is the platform shift designed to address all three at once.

In addition to these three core dimensions, Headless 360 also changes how governance is enforced and how you think about channel strategy. When the UI is no longer the primary surface, you can’t rely on a mix of UI constraints and schema rules for governance; instead, full schema-level enforcement (via validation rules, your sharing model, and Apex) is required. And rather than building per-channel UIs (for Lightning, mobile, and community, for example), with Headless 360 you define a capability once and let it render wherever it’s called, including Slack, ChatGPT, mobile, or web.

What Headless 360 actually is

Headless 360 makes every major Salesforce capability openly composable by agents, not just by code paths a human developer wrote in advance. The full surface, including 60+ MCP tools, 30+ coding skills, 4,000+ existing APIs, and 220+ CLI commands, is accessible to any authenticated caller that your trust layer authorizes.

This isn’t new territory for Salesforce developers. You’ve been building headless for years. This includes Experience Cloud sites with React frontends, Node.js apps over REST APIs, and mobile apps via the Mobile SDK. What’s new is that AI models can now dynamically discover, use, and compose capabilities at runtime, without custom integration code written in advance for each one. 

When we say “openly composable by agents” the openly matters. Openness is the default posture. Governance is what makes it safe.

Headless 360 in practice

The capabilities Headless 360 unlocks aren’t abstract. Here’s what they look like in practice. 

Build with any coding agent

The 60+ MCP tools and 30+ preconfigured coding skills give any compatible coding agent live access to your org. Agentforce Vibes 2.0 brings that natively inside Salesforce, but the MCP surface is the same regardless of which agent you choose.

Here is a typical session: your coding agent connects to a scratch org via MCP, reads the object model, scaffolds an Apex service class with governor-limit-aware test coverage, and pushes it through your DevOps pipeline. Same deployment gates, same quality checks, same review process as code written by hand. The pipeline doesn’t care who initiated it. It cares whether the gates pass.

The platform doesn’t privilege one agent over another. Your org’s capabilities are the constant. The tooling around it is your choice. For more on the MCP tools and how they work, see our post Introducing MCP Support Across Salesforce.

Render in any UI framework

Headless decouples capability from rendering. In practice, this means the platform no longer constrains your UI framework. A clear example of this is multi-framework support.

With multi-framework support, React developers can now build on Salesforce using the same patterns they already know, with no requirement to use LWC. A front-end engineer who’s never touched Salesforce can build a governed, data-connected application using familiar tools, because the platform adapts to their stack.

That hiring constraint where you needed someone who knew LWC or Aura? It’s no longer a hard constraint. The platform meets developers where they already are.

Deploy on any surface

The Headless Experience Layer is a new runtime that decouples capability definitions from their rendering surface. You define a UI fragment and the Headless Experience Layer (HXL) handles rendering it as a Slack block, a mobile card, a ChatGPT response, or a voice interaction. Business logic, data, and permissions stay separated from the screen, so you define intent once and render natively everywhere, including Slack, mobile, ChatGPT, Claude, Teams, or any MCP-compatible client.

This is where the two MCP lanes start to converge. The same openness that lets a coding agent build against your org will also let a business agent serve your customers from it. Today, the build-time surface is mature. The runtime surface already supports straightforward use cases, such as a support agent pulling a case summary inside a Slack thread. The broader vision is that the same capability can be delivered across surfaces such as voice interactions or partner mobile apps, and that ultimately capabilities can be made available on any surface.

Protect with built-in governance

Headless doesn’t mean ungoverned. Governance is what makes it safe to build with any coding agent, render in any UI framework, and deploy on any surface. 

When a coding agent pushes an Apex class, it goes through the same validation, tests, and deployment gates as code written by a human. When an agent queries customer data through MCP, it hits the same sharing rules and field-level security as a user in Lightning.

Four things are enforced on every API call, every MCP tool invocation, and every CLI command:

• Identity. The agent acts as a specific user, not an anonymous caller. Permissions are scoped to that identity.
• Access. Sharing rules, field-level security, and permission sets enforce what the caller can see and do.
• Invocation scope. The agent can invoke only those tools and actions that have been explicitly exposed.
• Governance. Validation rules, triggers, approval chains, and governor limits all fire regardless of entry point.

The platform is open about what can be accessed. It is not open about who can access it, or how. That’s what makes it safe for openness to be the default posture.

Why most agentic prototypes don’t reach production

If you’ve tried to ship an agent, you’ve likely encountered this problem. The demo works. The stakeholder asks a question that breaks it. For example:

• “Will it respect the sharing rule on this object?”
• “What happens if the user doesn’t have permission to update that field?”
• “How does it know which approval policy applies?”

The LLM alone is not enough. The hard part is everything around it: business context, identity-aware permissions, policy enforcement, and the operational guarantees that make a system safe to run autonomously.

That gap is exactly what the Salesforce platform fills. The context an agent needs didn’t show up because someone wrote a clever prompt. It accumulated over years, in your data model, your validation rules, your sharing architecture, your trigger framework. Headless 360 makes that platform accessible to agents through a governed surface. You don’t rebuild context, permissions, or policy for your agents. They inherit it.

Testing is part of this too. Testing agents is different from testing deterministic code. You’re evaluating whether an agent’s reasoning led to an acceptable outcome across a distribution of inputs. Session tracing, custom scoring evals, and the Agentforce Testing Center are how you build confidence that an agent does the right thing in production. The discipline is familiar if you’ve built test suites for complex Apex, but the techniques are new.

What comes next

Headless 360 ships with MCP tools for coding and business agents, coding skills, multi-framework support, and the Experience Layer. That’s its surface. What makes it work is everything underneath: the state, the governance, the accumulated business logic of a platform you already know.

On the Salesforce Developer Blog and YouTube, we’re going deep on every building block, including skills for coding agents, multi-framework walkthroughs, trust and governance deep dives, testing and observability patterns, and more. Follow along and tell us where to go deeper.

The platform is open. Build on it.

Related Reading

• The Future of the Salesforce Developer in the Agentic AI Era
• Introducing Salesforce Headless 360
• Agentforce Vibes 2.0
• Headless Experience Layer
• Salesforce Multi-Framework
• Agentforce Labs
• Open Source Agentforce Skills

About the Author

Rene Winkelmeyer leads the Developer Advocacy team at Salesforce. His team focuses on Agentforce, Data 360, and the Salesforce Platform. In his spare time, you might find him still coding on GitHub. Follow Rene on LinkedIn.

The post What Salesforce Headless 360 Means For Developers appeared first on Salesforce Developers Blog.

Picture your customer rolling out a Flexcard for quick account overviews. It’s looking great on desktop, but on mobile? Actions don’t fire, data doesn’t render, and users bounce in frustration. These aren’t rare moments; they’re everyday headaches that slow deployments, breaking trust. End-to-end (E2E) testing fixes this mess by mimicking real journeys, so you can catch issues early and deploy with confidence.

Why E2E testing matters for Omnistudio

So, why do we need to test the entire journey rather than just checking if a button works? Testing the full user journey is critical because Omniscripts, Flexcards, and Data Mappers interact across boundaries where unit tests cannot validate the complete process. E2E testing plays out the full story of what a real user does, from start to finish across your components, like a demo of an actual customer journey. 

E2E testing also helps you:

• Simulate full user journeys across Omnistudio components: E2E testing plays out that whole flow, just like a real person clicking through components, spotting weird hiccups along the way.
• Catch integration issues early for partners building solutions: E2E grabs those sneaky spots where Data Mappers don’t feed data right into an Omniscript, saving you from frantic fixes right before demo day.
• Reduce post-go-live bugs: E2E irons out production surprises, reducing post-deployment defect tickets for your teams, giving you greater confidence in deployments.
• Validate real-world flows to record creation: Unit tests poke single buttons (like does this button click? Does that data pull?). E2E runs the full race, like submitting a lead form, watching it zip through conditions, and landing as a shiny new record in Salesforce.
• Find the real problem: When tests flake, E2E pinpoints the issue past the pretty interface and discovers whether or not the backend logic is actually the part causing the crash.
• Confirm that the handoff works: Picture Omniscript passing data to a Flexcard. E2E checks that the pass doesn’t drop, enabling easy jumps between your UI pieces.
• Manage your own path: Users don’t always go straight. E2E tests the sunny “submit and succeed” path plus detours like errors or save-for-later resumes, covering all those branching stories.

High risk components to test first

When figuring out where to focus your testing energy, it helps to know which parts of the system are most likely to break. Based on what teams see in the real world, here are the high-risk areas you need to watch.

Omniscripts account for approximately 50% of reported Omnistudio UI bugs, making them the highest-priority target for E2E testing. Because this is the layer the user interacts with directly and is also most visible, it typically receives the highest number of reported bugs. They juggle steps, conditions, inputs, and data swaps with Data Mappers all in one go. Mess up here, and your whole customer journey, like a service request form, grinds to a halt.

Next is Flexcards because these cards are used to display critical data and handle user actions, and any glitch here is immediately obvious to the user. 

Finally, we have Data Mappers. If mappings fail (say, wrong field names or null values), Omniscripts starve for information, also breaking Flexcards downstream.

Note: All data presented above is derived from internal testing.

Tools for Omnistudio test automation: UTAM, Playwright, and Selenium

Automation sounds cool, but here’s the headache: Omnistudio pieces shift a lot. You might spend days building a perfect test script, and it runs beautifully, until the next release drops. Suddenly, everything breaks. Why? The main culprit here is dynamic element IDs. Because Omniscripts and Flexcards are built on Lightning Web Components (LWC), they sit inside a Shadow DOM. So your usual CSS or XPath selectors become fragile and start failing whenever Salesforce moves with seasonal updates. It breaks everything at the backend. One day your test clicks Submit just fine, and the next, it hunts for a button that moves and your whole pipeline stalls.

To solve this problem, Salesforce created the UI test automation model (UTAM), a proprietary tool designed specifically for LWC in Omnistudio. Omnistudio components are wrapped in Shadow Roots (that’s the Shadow DOM), which blocks standard Selenium selectors like CSS or XPath from reaching elements reliably, especially after Salesforce updates.

<!-- Source: https://mvnrepository.com/artifact/com.salesforce.utam/salesforce-pageobjects -->
<dependency>
    <groupId>com.salesforce.utam</groupId>
    <artifactId>salesforce-pageobjects</artifactId>
    <version>12.0.0</version>
    <scope>compile</scope>
</dependency>

Instead of relying on brittle selectors in scripts, UTAM uses JSON-based page objects. When a button’s location or structure changes in an Omnistudio update, you just update the JSON metadata once, and all your tests are instantly back on track. It creates a stable layer and lets scripts identify UI components consistently, no matter what changes underneath.

Choosing between UTAM, Playwright, and Selenium

Instead of relying on fragile identifiers, a smart move would be to grab the right tools that don’t flake. Use a combination of industry-standard tools and Salesforce-specific frameworks. Here’s when to use what:

• UTAM: Best for long-term maintenance and Salesforce-native components, giving you those stable JSON selectors that survive Omnistudio updates.
• Playwright: Playwright is best for speed and headless CI/CD pipelines. Fire it up without a visible browser to debug fast across browsers. 
• Cucumber with Selenium: Cucumber with Selenium is best for business stakeholder sign-off (BDD). You write tests in plain English and your team reads and approves it.

Launch with tweaks like test parameters (sandbox mode? Slow it down?). Then poke issues using element properties. Is it visible? Enabled? That’s how you spot and squash the real culprits without endless rewrites.

Testing scenarios by Omnistudio component

Now that we’ve covered the hurdles, let’s get practical: what exact scenarios should you click, check, and validate in your Omnistudio flows? Since Omnistudio is made up of different building blocks, we need to look at them individually. Here are the recommended testing scenarios to keep your application running smoothly.

These are the must-hit spots to make sure everything hangs together for real users.

Omniscript testing

These are your flow bosses, so test the moves that users make every day.

• Check step transitions & logic: Don’t just check if the Next button works. You need to validate the conditional visibility. This is the if/else logic, if a user selects Option A, does the form show them the correct follow-up question? If they select Option B, does it take them down the alternative path?
  
  
• Validation rules: Try to break it. Enter the wrong date format or leave a required field blank. Does the system stop you with a helpful error message (validation failure), or does it let you proceed and crash later?
  
  
• Save and resume: This is a big one. For long applications, users might not finish in one sitting. You must test the “Save for Later” functionality to ensure that their data is actually saved and can be resumed exactly where they left off without data loss.
  
  
• The ID trap: When automating these tests, avoid using dynamic IDs generated by the browser as these change frequently. Instead, use data-testid or stable CSS selectors. This keeps your test from breaking every time Salesforce updates the platform.
  
  

Flexcard testing

If Omniscript is the form, the Flexcard is the dashboard that needs to look sharp and act right. It displays data and offers quick actions.

• Validate data rendering: First, perform simple visual checks. Is the data showing up? Whether it is static info or contextual data pulled from the system, it needs to render correctly on the screen.
  
  
• Action buttons: Flexcards aren’t just for looking, they have buttons. Test these actions specifically. Does clicking the button successfully launch the correct Omniscript or navigate the user to the right page?
  
  
• Conditional states: Just like Omniscripts, Flexcards can change based on data. Test these conditional states, such as hover effects or expanded details, to ensure that the card looks different when it is supposed to.
  
  
• Selectors: Stick to application programming interface (API) name–based selectors for your automation to keep things stable. They stay reliable when Salesforce tweaks the UI.

Data Mapper testing

Now we are moving backstage. Data Mappers are responsible for moving data in and out of Salesforce. Even if the UI looks fine, a broken Data Mapper means the process fails.

• Test separately: You don’t always need the UI to test these. Isolate them to see if they work on their own since they’re data pipes.
  
  
• Input and mappings: Check the input payload correctness. If you send data in, does it map to the correct fields in the database?
  
  
• Edge cases: What happens if a field is missing or comes through as null? You need to test these edge cases to make sure that the Data Mapper doesn’t just crash when the data isn’t perfect.
  
  

Integration Procedures (IPs) testing 

Omnistudio’s Integration Procedures is like an engine room that handles multiple steps and actions at once.

• Orchestration sequence: Verify that the sequence fires in the right order without skips. Does the Integration Procedure call the Data Mapper, then the external API, and then send the email?
  
  
• Error branching and timeouts: It’s critical to test system failures.” What happens if an external service takes too long to respond (timeout handling)? Does the IP handle the error gracefully, or does the whole screen freeze?
• Mock services: In your lower testing environments (like sandbox), use mock services to simulate external responses. This lets you test the logic without relying on (or waiting for) real-world third-party systems.

Pretesting checklist for Omnistudio 

Before you even write your first test script or click Run, there are some best practices that you should keep in mind. Testing isn’t just about finding bugs; it’s about making sure that the design actually works for the human using it.

Here is a checklist of what to consider before you start testing.

Verify backend object updates

Don’t get too distracted by the flashy buttons just yet. Start with the win you want. Your primary goal is to verify that the business actually got what it needed. We call this testing the “happy path,” the core journey where everything goes right.

Ask yourself: When the user hits Submit, did the data actually land where it was supposed to? You need to confirm that the backend S-objects or business objects were updated correctly. If the form looks great but the database is empty, the test has failed.

Test dynamic UI state changes

Omnistudio interfaces are dynamic. They change based on who is looking at them. For example, a Flexcard might display different contextual data depending on the customer.

Before testing, verify that the UI visibility logic is sound. Does the screen state change correctly when new info is added? Are buttons going gray when they’re not ready? You need to ensure that the interface is displaying the right information at the right time.

Validate responsive UX

Step back and think about the user: does the flow feel natural on a phone or desktop? Test responsive tweaks and little touches like smooth transitions — no janky jumps that make users ditch the app mid-task. The Omniscript is the “face” of your application, and because it is the part that users touch, it usually receives the highest number of bug reports.

However, remember that a bad UX isn’t always a code error, sometimes it’s just a confusing design. When planning your tests, think about the flow. Is it intuitive? Since the root cause of a UI bug is often actually a backend configuration issue, ensuring a smooth UX means making sure that the backend supports the frontend easily.

Test error messaging paths

What happens when things go wrong? Your testing plan needs to cover more than just success stories. You must account for “unhappy paths,” such as validation failures where a user enters the wrong data.

You also need to verify how the system handles big failures, like an Integration Procedure timeout or an Apex exception. Does the user get a helpful error message, or does the screen just freeze? Check that success pop-ups say things like “Claim submitted, thanks!” and errors spell out next steps, like “Missing phone number, add it here,” instead of vague tech jargon.

Accessibility verification

Don’t leave anyone behind — this is non-negotiable. When designing and testing Omniscripts, you must consider accessibility because the UI needs to work for everyone.

Your tests should ensure that the application caters to all customers, including those who may be visually impaired. Test for screen readers announcing fields correctly, keyboard-only navigation through your Flexcards, and color contrasts that pop for all eyes.

Conditional navigation outcomes

Finally, where does the user go next? Omnistudio relies heavily on conditional logic, basically if/else statements that send users down different paths based on their choices.

You need to verify these alternative paths. If a user clicks Save for Later, does the system actually let them resume later without losing data? If they answer Yes to a specific question, does the navigation direct them to the correct follow-up screen? Make sure that every road leads to the right destination.

Next steps

There you have it. Tackling Omnistudio testing in the simplest way possible without it being a nightmare of flaky scripts and hidden bugs. Start with your highest-traffic Omniscript. Write one UTAM page object for its first step, validate a single field input and step transition, and run it against your sandbox. Once that passes reliably across two Salesforce releases, expand coverage to the scenarios listed above. Build rock-solid flows with clear messaging and confident deployments that your customers love.

Resources

Video: FlexCards and OmniScripts | Salesforce Industries 

Trailhead: Omnistudio Development Essentials 

Trailhead: Omnistudio Omniscript Fundamentals

Video: Build Low-Code Digital Experiences with All-New OmniStudio  | TDX Bengaluru 

About the authors

Anjali Gupta is a Senior Product Manager at Salesforce who is specialised in horizontal cloud and platform. She is focused on Omnistudio roadmap and strategies ensuring customer growth and success. Follow her on LinkedIn.

Payal Verma is a Content Marketing Analyst at Salesforce with over four years of professional experience in product, brand, and content marketing. She has worked across both B2B and B2C markets, developing and executing content strategies for a wide range of formats, including blogs, social media, scripts, eBooks, and guides. Follow her on LinkedIn.

The post Conduct End-to-End Testing for Salesforce Omnistudio-based UIs appeared first on Salesforce Developers Blog.

Custom Lightning types (CLTs) meet this need. They let you render Lightning Web Components (LWC) directly within Agentforce experiences in Lightning Experience, Enhanced Chat v2, and Experience Builder. You can collect structured input through a custom form and display results as rich, styled cards or in your own custom-designed user interface. CLTs are also the foundation of Headless Experience Layer, announced at TDX 2026, which extends rich agentic UI beyond Salesforce — to Slack, ChatGPT, and more. Define Once, Deploy Anywhere, Secure Everywhere — that’s the headless experience layer vision. Building with CLTs today means you’re already working with the primitives that power the next generation of multi-surface agentic experiences.

This post walks through how to work with CLTs in Agent Script to render a rich experience within Agentforce. All code examples come from the Agent Script Custom Lightning Type example recipe.

What are custom Lightning types in Agentforce?

Custom Lightning types create a bridge between your Agent Script, your actions backed by Apex or flows, and your LWC components. They override the default text-based UI with custom LWC components for both input and output.

Below is an example of custom Lightning types in action within the Agentforce experience.

The connection flows through three layers:

• Agent Script declares which action parameters need custom UI
• A Lightning Type Bundle maps each parameter type to an Apex data shape and an LWC component
• An LWC renders the user interface inside the conversation

This three-layer wiring enables you to keep a clean separation of concerns. Your Agent Script handles orchestration, Apex handles the data shape definition and the business logic, and LWC handles the final user experience.

The following diagram shows how a custom Lightning type is mapped with Agent Script and LWCs.

The following sections describe how to build and wire this connection step-by-step.

Step 1: Prepare and deploy your Apex and LWC 

Defining the Apex data shape

Each CLT needs an Apex class that defines the data structure. Use inner classes within your service class, with each field annotated with @InvocableVariable. The Lightning Type Bundle references these inner classes as the schema source.

public class CaseInput {
    @InvocableVariable(label='Subject' required=true)
    public String subject;

    @InvocableVariable(label='Priority')
    public String priority;

    @InvocableVariable(label='Description')
    public String description;
}

One important detail: the wrapper class field names must exactly match the action parameter names from your Agent Script. If your action input is called case_data, the request wrapper field must also be named case_data. A mismatch here silently breaks the entire CLT connection.

View the code: See the full CaseSubmissionService.cls for the complete Apex implementation including the request/response wrappers.

Building the editor and renderer LWC

The editor LWC collects user input inside the chat. Two things make it work with Agentforce: the meta XML must target lightning__AgentforceInput with a targetType matching the CLT name, and the component must dispatch a valuechange event whenever the user modifies the form.

 <target>lightning__AgentforceInput</target>
 <targetType name="c__caseInput" />

The valuechange event payload field names must match the Apex CaseInput class fields. The platform listens for this event to capture and forward the form data to your Apex action.

Here is the simple example snippet showing the valuechange event in practice:

import { LightningElement } from 'lwc';

export default class CustomCaseForm extends LightningElement {
    // Local state variables for the form
    subject = '';
    priority = 'Medium';
    description = '';

    // Handler triggered when the user types in the Subject input
    handleSubjectChange(event) {
        // 1. Update the specific field that changed
        this.subject = event.target.value;

        // 2. Dispatch the FULL state of the form
        this.dispatchFormState();
    }

    // You would have similar handlers for Priority and Description...

    // Centralized method to dispatch the complete current state
    dispatchFormState() {
        this.dispatchEvent(
            new CustomEvent('valuechange', {
                detail: {
                    value: {
                        // ALWAYS include all fields to match the Apex CaseInput class
                        subject: this.subject,
                        priority: this.priority,
                        description: this.description
                    }
                }
            })
        );
    }
}

The renderer LWC displays the result card after the action completes. It targets lightning__AgentforceOutput with a sourceType (not targetType). The component receives the full result object via @api value and uses getters to render the fields with Salesforce Lightning Design System styling.

<target>lightning__AgentforceOutput</target>
<sourceType name="c__caseResult" />

View the code: See the full caseInputEditor and caseResultRenderer components in the repo.

Wiring Lightning type bundles

The Lightning Type Bundle is a folder containing JSON configuration files. The schema.json file points to your Apex inner class using $ syntax. For example, c__CaseSubmissionService$CaseInput references the CaseInput inner class.

{
    "title": "Case Input",
    "lightning:type": "@apexClassType/c__CaseSubmissionService$CaseInput"
}

A second file tells the platform which LWC component to render. Input types use an editor.json file. Output types use a renderer.json file. The "$" root key means “override the UI for the entire type”.

{
    "editor": {
        "componentOverrides": {
            "$": { "definition": "c/caseInput" }
        }
    }
}

View the code: See the complete Lightning Type Bundle files in the repository

Step 2: Create the agent action asset

The GenAiFunction is the metadata layer that registers your action with Agentforce. Without it, Agentforce cannot resolve your CLT bindings. Every CLT-based action needs a corresponding GenAiFunction in the genAiFunctions folder.

The function metadata XML points to your Apex invocable class:

<GenAiFunction xmlns="http://soap.sforce.com/2006/04/metadata">
    <developerName>Submit_Case</developerName>
    <invocationTarget>CaseSubmissionService</invocationTarget>
    <invocationTargetType>apex</invocationTargetType>
    <masterLabel>Submit Case</masterLabel>
</GenAiFunction>

Alongside this XML, you define input and output schema JSON files. These schemas declare the lightning:type for each parameter, which is how the platform knows to look up your Lightning Type Bundle.

The input schema marks case_data as a user input with the CLT type c__caseInput:

{
    "properties": {
        "case_data": {
            "lightning:type": "c__caseInput",
            "copilotAction:isUserInput": true
        }
    }
}

The output schema marks case_result as displayable with the CLT type c__caseResult:

{
    "properties": {
        "case_result": {
            "lightning:type": "c__caseResult",
            "copilotAction:isDisplayable": true
        }
    }
}

See the full GenAI Function files in the repository.

You can build the Agent action metadata directly in the Salesforce user interface. 

To do this, navigate from Setup to Agentforce Assets, then select Actions and create a new agent action by following below steps.

1. Setup → Agentforce Assets → Click on Actions → Click on New Agent Action
2. Fill in the agent action details like Inputs and Outputs description and data types.
3. Ensure you are setting up the Input Rendering and Output Rendering with the right custom Lightning types

Step 3: Define the action in Agent Script

While the following sections cover action definition via metadata in detail, the Agentforce Builder UI enforces proper naming constraints automatically and is the recommended starting point for most teams.

Everything hinges on the action definition in your Agent Script. Two properties control CLT rendering: complex_data_type_name links a parameter to a Lightning type, and is_user_input or is_displayable tells the platform to render the custom component.

actions:
   submit_case:
      inputs:
         case_data: object
            is_user_input: True
            complex_data_type_name: "c__caseInput"
      outputs:
         case_result: object
            complex_data_type_name: "c__caseResult"
            is_displayable: True
      target: "apex://CaseSubmissionService"
      source: "Submit_Case"

Notice the source property at the bottom. This is a critical piece that connects Agent Script to an Agentforce action (represented as GenAiFunction in the metadata as discussed in step 2). The value “Submit_Case" must match the developer name of a GenAiFunction metadata file in your project.

Triggering the CLT editor in Agent Script

One detail that is easy to miss: your Agent Script subagent instructions must include the user_input keyword when calling the action. This tells the platform to render the CLT editor component instead of collecting input through plain text.

reasoning:
   instructions: ->
      | Call {!@actions.submit_case} action's user_input tool to collect the case details.

Without this phrase, the platform falls back to text-based input collection. The user_input keyword is what activates the entire CLT editor rendering pipeline.

Important note: In Agent Script the subagent’s action name and the subagent reasoning action name should be exactly the same.

View the code: See the full instructions for the Agent Script in the repository.

The alignment checklist

CLTs require precise naming across multiple files. If any link in the chain breaks, the platform silently falls back to the default text UI with no error message. The following table shows what must stay aligned:

Check these connections first when debugging. The silent fallback behavior makes misalignment hard to spot.

Conclusion

Custom Lightning types bring the power of Lightning Web Components into Agentforce conversations. You can collect precise, validated data through custom forms and present results in rich, branded cards, all without leaving the Agentforce experience.

The key pieces are the Agent Script action definition and the action in the Agentforce assets, the Lightning Type Bundle, and the LWC components. Each layer has a specific job, and the naming alignment across all four layers is what makes the whole system work.

This pattern opens up new possibilities for building enterprise-grade agent experiences. Forms with dropdowns and validation, structured result cards with conditional styling, and multistep workflows can all be rendered inline within the conversation.

The complete working example is available in the Agent Script Recipes repository under the customLightningTypes recipe. Clone the repo, deploy to a scratch org or Developer Edition org, and start building your own CLT-powered Agentforce agents.

Resources

• Lightning Types Developer Guide
• Agent Script Recipes Sample App
• Agent Script Documentation

About the authors

Mohith Shrivastava is a Principal Developer Advocate at Salesforce with 15 years of experience building enterprise-scale products on the Agentforce 360 Platform. Mohith is currently among the lead contributors on Salesforce Stack Exchange, a developer forum where Salesforce Developers can ask questions and share knowledge. You can follow him on LinkedIn.

Parvinder Singh is a Senior Forward Deployed Engineer at Salesforce with deep expertise in Agentforce and Agent Script, specializing in building intelligent, customized agent experiences for enterprise customers. He is passionate about sharing developer insights with the community and bringing practical, field-tested patterns to life through hands-on content.

The post Use Custom Lightning Types in Agent Script for Rich Agent UI appeared first on Salesforce Developers Blog.