As a Salesforce developer, you know that a new building experience usually means changes under the hood. To support this faster, safer way of working, the underlying metadata had to evolve. This post walks through the new developer workflow, why the metadata changed, and how to structure your packages for deployment.

The new authoring workflow: from script to metadata

To understand the metadata changes, it helps to understand the new mental model for building agents. The new Agentforce Builder introduces a clear separation between your “intent” (what you want the agent to do) and the engine’s “execution” (the active metadata).

Here is how the lifecycle works:

1. Write and build: You author your agent using the Agentforce Builder UI or directly via an Agent Script. This human-readable script is saved as a single .agent file within a new AiAuthoringBundle metadata type, representing your agent’s configuration. At this stage, your work is a draft.
2. Iterate and preview: You test your draft using preview modes (like Simulation mode, which safely isolates your tests from org data).
3. Publish: When you are happy with the draft, you commit your version. This is the crucial moment. It is only during commit that the platform translates your .agent script into the active Bot and GenAiPlannerBundle metadata that is actually used to run your agent.

Why the metadata changed: bundling local assets

So, why did we change the GenAiPlannerBundle metadata structure for this new builder? The answer comes down to isolation and deployment reliability.

Historically, Agentforce assets were global. If you modified a shared subagent for one agent, you triggered a “global ripple” that instantly impacted every other agent using that subagent. To solve this, Salesforce introduced Local Assets (including local subagents and local actions), which clone a global asset and anchor it specifically to one agent version. Once local, you can edit it safely without breaking anything else.

While Local Assets have been available for a while, the GenAiPlannerBundle metadata needed to be updated to fully support them in CI/CD pipelines. Previously, agent metadata was distributed across metadata types. The main change in the new GenAiPlannerBundle is bundleization. In short, we restructured the metadata to ensure that all dependent agent assets related to any given version are deployed and available in one centralized folder structure. This prevents deployment errors and ensures your agent has everything it needs to function exactly as it did in preview.

Under the hood: the new metadata structure

Instead of scattered files, the directory structure now cleanly encapsulates everything specific to each agent version.

AiAuthoringBundles is a directory that contains all the design-time artifacts for the agent. The bundle supports multiple versioned subdirectories (e.g., AgentName_1, AgentName_2, etc.), each with its own .agent file and .bundle-meta.xml descriptor. These correspond to different lifecycle states such as draft and committed versions of the agent. Within each version:

• The .agent file holds the agent’s configuration written in Agent Script: name, label, description, system instructions, subagent definitions, reasoning instructions, variables, conditionals, and tool/action references.
• The metadata descriptor (.bundle-meta.xml) declares the bundleType (e.g., AGENT) and the target agent version it points to.

Bot contains information common to all versions of the agent, such as the agent type and useful settings like whether to log private data. It also holds agent-level context variables, which are distinct from the conversation variables that live inside each BotVersion. BotVersion details include conversation variables, welcome message, and transfer conversation message.

GenAiPlannerBundle contains all data relevant to each published version of an agent. Within each version you can find:

• .genAiPlannerBundle, which is the main metadata file containing the agent’s runtime configuration: subagents, instructions, actions, and orchestration logic.
• agentGraph, a .json file encoding all transitions and conditions between subagents.
• agentScript, an encoded copy of the version agent script. This is a snapshot kept for sync-detection purposes, so Agent Builder can tell the user if the script is out of sync with the published/committed metadata.
• plannerActions, agent-level actions (not tied to any specific subagent).
• localActions, subagent-scoped actions, nested under their subagent folder. The input/schema.json and output/schema.json files define the typed parameters for each action.

Note: GenAiPlannerBundle versions are auto-incremented in commit order, so they won’t necessarily match AiAuthoringBundle version numbers.

Remember that both Bot and GenAiPlannerBundle metadata are created automatically when a version is committed, and AiAuthoringBundle, agentScript, and agentGraph are only present on the new Agentforce Builder agent metadata.

Example: GenAiPlannerBundle XML

When you look at the XML, you’ll see it now clearly reflects these localized subagents and actions:

<GenAiPlannerBundle>
   <description>New agent description</description>
   <localTopicLinks>
       <genAiPluginName>Retrieve_data_from_the_Knowledge_Base_16jxx0000001234</genAiPluginName>
   </localTopicLinks>
   <localTopics>
       <fullName>Retrieve_data_from_the_Knowledge_Base_16jxx0000001234</fullName>
       <description>Use the retriever action</description>
       <pluginType>Topic</pluginType>
       <masterLabel>Retrieve data from Knowledge Base</masterLabel>
       <genAiPluginInstructions>
           <description>You are an AI Agent.</description>
           <sortOrder>1</sortOrder>
       </genAiPluginInstructions>
       <localActionLinks>
            <functionName>File_test_retriever_179xx0000001234</functionName>
       </localActionLinks>
       <localActions>
           <fullName>File_test_retriever_179xx0000001234</fullName>
           <invocationTarget>File_test_retriever_1234</invocationTarget>
           <invocationTargetType>apex:testRetriever</invocationTargetType>
           <source>File_test_retriever</source>
       </localActions>
   </localTopics>
   <plannerType>AiCopilot__ReAct</plannerType>
</GenAiPlannerBundle>

Each <localTopics> block defines a subagent’s full configuration: its unique API name, description, plugin type, deterministically ordered instructions, and all the local actions available to it. Optionally, a reference to the source subagent may also be present.

Each <localActions> block defines the action’s name, master label, invocation target and type, and optionally a source reference to the global action it was cloned from. Additional UX-related settings, such as confirmation behavior and loading text configuration, may also be included.

Sample package.xml with metadata of new model

To successfully retrieve and deploy your agents built with the new Agentforce Builder, ensure your package.xml includes the new bundle types. Below is a sample manifest designed for the Spring ’26 metadata structure.

<?xml version="1.0" encoding="UTF-8"?>
<Package xmlns="http://soap.sforce.com/2006/04/metadata">
    <types>
        <members>*</members>
        <name>GenAiFunction</name>
    </types>
    <types>
        <members>*</members>
        <name>Bot</name>
    </types>
    <types>
        <members>*</members>
        <name>GenAiPlannerBundle</name>
    </types>
    <types>
        <members>*</members>
        <name>GenAiPlugin</name>
    </types>
    <types>
        <members>*</members>
        <name>AiAuthoringBundle</name>
    </types>
    <version>66.0</version>
</Package>

Note: This manifest uses API version 66.0, which is required to support GenAiPlannerBundle and the new agent metadata types in this example.

Developer tips for the new Agentforce Builder

Here are two ways to take advantage of development lifecycle improvements enabled by this new architecture:

• Use human-readable script files to streamline code reviews. In the legacy Agentforce Builder world, reviewing agent changes meant comparing complex, scattered XML files in your pull requests. Now, you can compare the human-readable .agent script files. This makes peer reviews significantly faster and easier to understand.
• Use AiAuthoringBundle to simplify CI/CD. For most pipelines, you can keep just the AiAuthoringBundle in your source control. When you need the agent in a new org, you deploy it, version commit to generate the runtime metadata, and finally activate it. This can be done automatically using the following commands:sf project deploy start --metadata AiAuthoringBundle --target-org my-orgsf agent publish authoring-bundle --api-name MyAuthoringBundle --target-org my-org

sf agent activate --api-name MyAuthoringBundle --version 2 --target-org my-orgIf your team wants a fully automated pipeline with no manual steps, you can maintain both AiAuthoringBundle and GenAiPlannerBundle in your source control and deploy them simultaneously, achieving a committed state. The system recognizes the incoming script alongside its pre-realized metadata and skips the manual commit step entirely, so your agent is live immediately after deployment.sf project deploy start --metadata AiAuthoringBundle,GenAiPlannerBundle --target-org my-org

Conclusion

The bundleized metadata structure is what makes the isolated, version-safe authoring possible in the new Agentforce Builder. With it, your metadata is cleaner, your code reviews are easier, and your deployments are more reliable. 

Get started with Agent Script and the new Agentforce Builder by completing this Trailhead module, reading more on our blog, watching our Agent Script decoded playlist on YouTube, or registering for one of the upcoming Agentforce Builder workshops.

Resources

• Release Notes:
  • Versioning and Editing Agents
  • Test and Customize your Agents with Improved Agent Versions
• Help Docs:
  • Editing Standard Agent Action Reference Actions 
• Packaging:
  • Package Agentforce Metadata Components
  • Develop and Package Agent Templates Using Scratch Orgs 
• Metadata updates:
  • GenAiPlannerBundle
  • GenAiFunction
  • GenAiPlugin
  • GenAiPluginInstructionDef

Shreyas is a Lead Software Engineer at Salesforce AI in San Francisco and a Fellow of the Institute of Analytics (FIoA). As a lead member of the Agentforce Core Platform team, he co-designed and led the Agentforce Versioning System and Local Assets Configuration changes to bring enterprise-grade ALM to Agentforce. You can follow him on Linkedin.

Alba Rivas works as a Principal Developer Advocate at Salesforce. You can follow her on Linkedin.

The post The New Agentforce Metadata and Development Lifecycle appeared first on Salesforce Developers Blog.

Headless 360 and Salesforce Hosted MCP Servers introduce a shift from UI-first to agent-first. Every capability on the platform — data access, business logic, automation — becomes consumable through CLIs, APIs, and the Model Context Protocol (MCP), an open standard that lets AI agents discover and call tools hosted on external servers. AI agents running in Slack, Claude, ChatGPT, Cursor, or any MCP-compatible client can now reach into Salesforce and invoke your business logic directly. No login screen. No navigation. No UI at all.

In this blog post, we’ll show you how to write custom Apex business logic, expose it as a Hosted MCP Server on Salesforce, deploy the necessary metadata to configure authentication, and connect an external AI agent like Claude to invoke your logic.

What makes Salesforce Hosted MCP different

A Salesforce Hosted MCP Server is a platform-managed endpoint that exposes Apex actions, Apex REST, and external APIs as discoverable tools for AI agents through the Model Context Protocol. Unlike self-hosted MCP servers, it runs on Salesforce infrastructure handling authentication, tool discovery, and request routing without external deployment.

A Salesforce Hosted MCP tool delivers three things that no raw database can match out-of-the-box: 

• Pre-built business object graphs (e.g., Account → Opportunity → Product → Revenue) with relationships already modeled 
• A record-level sharing model that automatically scopes results to what this specific caller is allowed to see 
• Decades of accumulated business process metadata that each customer’s org has already encoded over years of operation

Example: Pipeline intelligence

The example exposes a pipeline intelligence tool built in Apex. An AI agent passes an account name, and the tool returns a complete analysis: open deals with risk levels, weighted pipeline value, product line items, and an overall health assessment.

The Apex @InvocableMethod (see docs) that powers flows and Agentforce actions now also serves external AI agents through MCP. Write the logic once; it runs everywhere.

The Apex class uses global access (required for MCP tool discovery) and structures the method as a deep module, absorbing interpretive logic so that the Agent receives insight, not raw data.

The method starts by resolving the account using fuzzy matching: trying an exact name match first, then falling back to a LIKE pattern.

global with sharing class AccountPipelineService {
    @InvocableMethod(
        label='Get Account Pipeline'
        description='Returns open opportunities with line items and computed business intelligence (risk levels, weighted pipeline, health summary) for a given account name. Supports fuzzy matching.'
    )
global static List getAccountPipeline(List requests) {
  // All private methods is invoked here..
}
private static Account resolveAccount(String accountName, PipelineResult result) {
     List accounts = [
         SELECT Id, Name, Industry, AnnualRevenue
         FROM Account
         WHERE Name = :accountName
         WITH USER_MODE
         LIMIT 1
     ];

     if (!accounts.isEmpty()) {
         result.matchType = 'exact';
         return accounts[0];
     }

     String fuzzyPattern = '%' + accountName + '%';
     accounts = [
         SELECT Id, Name, Industry, AnnualRevenue
         FROM Account
         WHERE Name LIKE :fuzzyPattern
         WITH USER_MODE
         LIMIT 1
     ];

     if (!accounts.isEmpty()) {
         result.matchType = 'partial';
         return accounts[0];
     }

     result.accountFound = false;
     result.message = 'No account found matching: ' + accountName;
     return null;
 }

}

When no account matches, the tool returns a structured response with accountFound: false and a descriptive message, giving the AI agent a clear signal to ask the user for clarification rather than failing silently.

Once the account is resolved, a single SOQL query traverses the full business object graph.

List opportunities = [
     SELECT Id, Name, Amount, StageName, Probability, CloseDate,
         Owner.Name,
         (SELECT Id, Product2.Name, Quantity, UnitPrice, TotalPrice
          FROM OpportunityLineItems)
     FROM Opportunity
     WHERE AccountId = :acc.Id
         AND IsClosed = false
     WITH USER_MODE
     ORDER BY Amount DESC NULLS LAST
 ];

Each deal is then scored for risk based on probability and proximity to close date.

private static String assessRisk(Decimal probability, Integer daysUntilClose) {
     if (daysUntilClose <= 30 && (probability == null || probability < 60)) {
         return 'High';
     }
     if (probability == null || probability < 50) {
         return 'Medium';
     }
     return 'Low';
 }

The overall pipeline health is a multi-factor assessment combining high-risk ratio and weighted conversion strength.

private static String assessPipelineHealth(
     Integer totalDeals, Integer highRiskCount,
     Decimal weightedValue, Decimal totalValue
 ) {
     if (totalDeals == 0) {
         return 'Empty — no open deals';
     }
     Decimal highRiskRatio = (Decimal) highRiskCount / totalDeals;
     Decimal conversionStrength = totalValue > 0 ? weightedValue / totalValue : 0;

     if (highRiskRatio > 0.5) {
         return 'Weak — majority of deals are high risk';
     }
     if (conversionStrength < 0.4) {
         return 'At Risk — low weighted conversion across pipeline';
     }
     if (highRiskRatio <= 0.2 && conversionStrength >= 0.6) {
         return 'Strong — high confidence, low risk';
     }
     return 'Moderate — mixed signals, review recommended';
 }

Key design decisions:

• Fuzzy matching: The tool resolves partial account names (“Acme” finds “Acme Corporation”), reducing errors for AI agents that may not know exact names.
• Computed fields: riskLevel, weightedPipelineValue, and pipelineHealthSummary are calculated inside the method. The AI agent never needs to interpret raw probabilities or close dates.
• WITH USER_MODE: SOQL queries enforce the running user’s field-level and object-level security automatically.
• Cross-object traversal: This is a single tool call that spans Account, Opportunity, OpportunityLineItem, Product2, and User.

Registering the Hosted MCP Server

With the Apex class deployed, the next step is the MCP Server Definition, the metadata that maps your Apex action to a named tool and controls how AI agents discover it.

Below is an example deployable metadata for registering the hosted MCP server. The metadata can be source controlled and CI/CD deployed alongside the Apex class.

<?xml version="1.0" encoding="UTF-8"?>
<McpServerDefinition xmlns="http://soap.sforce.com/2006/04/metadata">
    <description>Analyzes account pipeline health with deal risk scoring
    and weighted forecasts by traversing Salesforce's native business
    object graph.</description>
    <masterLabel>Pipeline Intelligence</masterLabel>
    <tools>
        <apiDefinition>
            <apiIdentifier>aa:apex-AccountPipelineService</apiIdentifier>
            <apiSource>API_CATALOG</apiSource>
            <operation>AccountPipelineService</operation>
        </apiDefinition>
        <descriptionOverride>Returns open opportunities with line items
        and computed business intelligence (risk levels, weighted pipeline,
        health summary) for a given account name.
        Supports fuzzy matching.</descriptionOverride>
        <toolName>getAccountPipeline</toolName>
        <toolTitle>Get Account Pipeline</toolTitle>
    </tools>
</McpServerDefinition>

Note: MCP Registration can also be done via Salesforce User Interface. The steps are documented in the official documentation.

Why every field here matters

AI agents don’t read your code. When an MCP client calls tools/list, the agent sees exactly four things per tool: a name, a description, an inputSchema, and an outputSchema. That’s the entire basis for deciding whether to call your tool, what to pass, and how to interpret the response.

Each metadata field controls part of that picture.

From Apex annotations to agent-readable schemas

The apiDefinition is the bridge. It references the API Catalog entry auto-created from your global class. Salesforce reads the @InvocableVariable annotations on your input and output classes and generates the inputSchema and outputSchema that agents see.

// Input class → becomes inputSchema (wrapped in an inputs[] array)
 @InvocableVariable(required=true description='The name of the Account to look up pipeline for. Supports partial matching.')
 global String accountName;

 // Output class → becomes outputSchema (nested inside a response envelope as outputValues)
 @InvocableVariable(description='Overall pipeline health: Strong, Moderate, At Risk, or Weak')
 global String pipelineHealthSummary;

 @InvocableVariable(description='Pipeline value weighted by probability — realistic expected revenue')
 global Decimal weightedPipelineValue;

The description text you write in each @InvocableVariable annotation becomes the documentation that the agent reads. Without it, the agent sees field types but not meaning. Write them with the agent as your audience for example:

• Input fields: What to provide. “Supports partial matching” tells the agent it doesn’t need an exact name.
• Output fields: How to interpret. “Negative means overdue” on daysUntilClose prevents misreading.
• Enumerated values: List them. “Strong, Moderate, At Risk, or Weak” — no guessing.

Writing an effective descriptionOverride

The descriptionOverride drives tool selection for the Agent. A good description answers three questions:

1. What does it return?  “Returns open opportunities with line items and computed business intelligence (risk levels, weighted pipeline, health summary)”
2. What input does it need?  “For a given account name”
3. Any behavioral nuances?  “Supports fuzzy matching”

Anti-patterns to avoid:

1. Too vague: A description like “Gets pipeline data” gives the agent no signal about what makes this tool different from a generic query tool. The agent is unlikely to select it when multiple tools are available
2. Too implementation-focused: A description like “Executes SOQL against Opportunity with USER_MODE” tells the agent how the tool is built rather than what it delivers. The agent needs to know what business value the tool provides, not the internal mechanics.
3. Missing input guidance: If the description does not mention that the tool expects an account name, the agent has no way to know whether it should pass an account ID, an opportunity name, or something else entirely.

Activate the MCP Servers

After deploying, verify the server appears in Setup → MCP Servers and activate it. The server must be active before any MCP client can connect.

Setting up the External Client App

For an MCP client like Claude to authenticate, an External Client App with OAuth is required. This is configured through Setup as follows:

1. Navigate to Setup → External Client App Manager → New External Client App
2. Provide a label, description, and contact email
3. Enable OAuth with the following settings:

Once saved, the Consumer Key becomes available under Settings → Consumer Key and Secret. This is the client identifier that Claude uses to initiate the OAuth flow.

Note: External Client Apps can take up to 30 minutes to become fully active after creation. If Claude’s connection fails immediately after setup, wait and retry.

Connecting Claude

Connecting Claude to a Hosted MCP Server requires a server URL and an OAuth Client ID — nothing else.

1. In Claude.ai on the web, navigate to Settings → Connectors → Add custom connector.Note this assumes you have a Claude account and you are signed in. If you are signed in you can also directly add a connector.
2. Copy the Server URL from Salesforce Setup:

• Navigate to Setup → Integrations → MCP Servers
• Click the MCP Server name (e.g., “Pipeline Intelligence”)
• Under Authentication Details, copy the Server URL

The URL follows the pattern below:

For this project, the URL is https://api.salesforce.com/platform/mcp/v1/sandbox/Pipeline_Intelligence

Note: You can also copy this from the user interface by navigating to the following: 

1. Set the Server URL in Claude by pasting the URL from Step 2
2. Paste the OAuth Consumer Key from the External Client App settings
3. Click Connect and authenticate with the Salesforce org credentials

Once connected, Claude can invoke the tool naturally:

“What’s happening with Acme’s pipeline?”

The response includes the matched account, open deals with per-deal risk levels, weighted pipeline value, product line items, and an overall health assessment — all computed by the Apex running on Salesforce, all respecting the user’s security context.

Note: If you are having issues connecting to MCP servers refer to the troubleshooting section in the official docs.

The Headless 360 shift

For Apex developers, Headless 360 is a fundamental shift in how you design code. The consumer is no longer a human clicking through a UI; it’s an AI agent that reads tool descriptions, passes structured inputs, and reasons over structured outputs.

Designing for agents means returning computed insight (not raw fields), writing precise descriptions (not UI labels), and building tools that are self-contained (not steps in a wizard).

For AI developers exploring MCP, Salesforce is a uniquely deep server: a pre-modeled business graph, governed security, and computed intelligence — not just another database behind an API.

A note on tool governance

Too many MCP tools overwhelm AI agents, they struggle to pick the right one, leading to irrelevant calls or confused responses. As more teams expose Apex logic as MCP tools, governance becomes important. The Salesforce Hosted MCP Servers Best Practices guide recommends keeping tools self-contained (each returns a useful result on its own), avoiding both over-granular tools that require prescribed sequences and over-broad tools that bundle unrelated operations. Give each tool

Get started

The complete source code for this blog post, setup scripts, and sample data are available on GitHub: headless-apex-mcp-tool. 

While this post demonstrates the connection with Claude, MCP is an open protocol. Any MCP-compatible agent — Slack, ChatGPT, or your own custom agent — can connect to the same Hosted MCP Server using the same URL and OAuth flow. Check the resources below to learn more.

Resources

• Connect MCP Clients
• External Client App Setup for MCP
• Connecting Claude to Salesforce MCP
• Hosted MCP Servers Developer Guide
• GitHub: Headless Apex MCP Tool

About the author

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.

The post Expose Custom Apex as a Hosted MCP Tool for Agents appeared first on Salesforce Developers Blog.

I built AgentLens when I was writing Agent Script recipes, the declarative definitions that control how Agentforce route, respond, and invoke tools. I needed a way to quickly debug what my agents were doing during execution in order to build a clear mental model of their orchestration. Whether I was reviewing agent behavior, or instructing a coding agent like Agentforce Vibes or Claude Code to modify Agent Script, I needed to understand the flow. Which subagents talked to each other? What does the state machine look like inside each one of these subagents? Where do handoffs happen, and why?.

Agentforce already provides a powerful agent platform tracing capabilities. Every execution is captured — subagent routing, LLM calls, tool invocations, variable mutations, all of it. That trace data is rich and complete, but what was missing for me was a fast, intuitive way to visualize it. I wanted a tool that would allow me to build a mental model in minutes, so I could describe orchestration clearly to a coding agent, judge whether its output was correct, or spot patterns in how my Agentforce agents behave.

That’s why I built AgentLens.

How Agentforce orchestration works (and why it matters for debugging)

Before diving into the tool, it helps to understand the two core structures that define how an Agentforce agent executes.

An Agentforce agent is composed of subagents, and each is responsible for a specific capability, such as order lookup, case routing, or knowledge retrieval for an order management agent. When a user sends a message, the orchestrator decides which subagent should handle the request and hands off the conversation. That subagent may invoke another in turn. The result is a directed acyclic graph (DAG), a network of subagents connected by handoff edges, where each edge represents one subagent passing control to another. Understanding this DAG tells you which subagent participated in a conversation and in what order.

Inside each subagent, execution follows a finite-state machine (FSM), a sequence of states connected by transitions. The subagent moves from state to state; it might start by calling the LLM, transition to a tool execution based on the LLM’s decision, then transition to a response or hand off to another subagent. Each state has defined transitions that determine what happens next. Understanding the FSM helps you see not just what a subagent did, but how it made decisions at each step: which tools it invoked, why transitions occurred, and where execution paths diverged.

When you’re debugging an Agentforce agent — or prompting a coding agent to fix one — you’re ultimately trying to answer two questions: 

1. “Did the right subagents get involved, in the right order?” (the DAG) 
2. “Did each subagent behave correctly internally?” (the FSM) 

The first tells you how the orchestration flowed across the system. The second tells you how decisions were made within each subagent. AgentLens makes both visible.

Visualize agent orchestration at three levels

AgentLens takes a single Agentforce trace JSON file and presents three connected views, each one level deeper than the last.

Agent Graph (the DAG): See which subagents communicated and in what order. Each node is a subagent, and each edge is a handoff. The graph shows you the full routing path that the conversation took, making it easy to spot unexpected routing, circular handoffs, or subagents that never got invoked when they should have been.

Finite-State Machine Diagram: Pick any subagent and see its internal execution as an interactive FSM. Nodes are color-coded: blue for LLM calls, teal for tool executions, green for responses. Each arrow is a state transition. Backward arrows reveal retries and loops — places where the subagent repeated a step, usually because a tool returned unexpected results or the LLM needed another pass. This is what tells you how a subagent made its decisions, and it’s what you describe to a coding agent when you want it to change the behavior.

In one of my recipes, the FSM showed the Routing subagent handing off to Fulfillment four times in a loop. The backward arrow made the pattern obvious: the handoff condition was matching too broadly. I described the FSM to Agentforce Vibes, and it tightened the guard clause in seconds.

Step-by-Step Trace Inspector: Walk through every agent execution event in sequence: the exact system prompt the model received, what it responded to, tool inputs and outputs, and variable diffs showing what changed. Use the arrow keys to navigate and handoffs automatically jump to the next agent, so you never lose the thread.

Filter controls let you toggle the visibility of variable updates, reasoning steps, enabled tools, and internal system variables. Show only what matters for the question you’re asking.

Debug Agent Script in VS Code or your browser

Analyzing Agentforce execution traces is the best way to understand its internal decision-making process. Let’s look at how you can easily visualize and debug these traces with AgentLens.

VS Code Extension: Search for AgentLens Visualizer in the Extensions panel or install it from the VS Code Marketplace. Once installed, right-click any trace JSON and select Open with AgentLens. You can still search for AgentLens Visualizer in Cursor or other IDEs, since the extension is available on the Open VSX Registry.

If your traces live at .sfdx/agents/**/traces/*.json (the default location used by the Salesforce CLI), the extension automatically adds an inline Open in AgentLens CodeLens at the top of the file. AgentLens also integrates with your VS Code theme, with support for light, dark, and high-contrast modes.

Web App: Head to msrivastav13.github.io/AgentLens, paste or upload a trace, and you’ll immediately see the agent graph — no install needed. 

For full offline use, download the index.html from the GitHub repo, open it in your browser, and paste the trace JSON from Agentforce Builder. Everything runs locally on your machine, and no trace data leaves your environment.

Where do Agentforce traces come from?

You can export traces from several places in the Agentforce ecosystem:

• SF CLI: Run sf agent preview -o <org>. Traces save to .sfdx/agents/ in your project.
• Agentforce DX Extension: Copy the plan response JSON directly from the trace viewer.
• Agentforce Builder: Run a conversation in the preview panel, then copy the trace JSON from conversation details.

The screenshot below shows how to copy trace JSON from the Agentforce Builder trace panel.

Once you’ve copied the trace, simply paste it into the AgentLens user interface to visualize the agent graph and step through the execution trace interactively. The screenshot below shows how AgentLens transforms the raw JSON copied from Agentforce Builder into an interactive debugging experience.

Development versus production: Pick the right tool

AgentLens is designed for the development loop, when you’re iterating on Agent Script, testing behavior in preview, and refining how your agent orchestrates. It helps you understand what happened in a test run, so you can make the next change with confidence.

For production observability, monitoring live agent conversations, tracking performance, and diagnosing issues in deployed agents, use Agent Platform Tracing. That’s the built-in platform capability for tracing agent execution at scale in your org.

A useful way to think about it:

• AgentLens is the workbench microscope you use while building
• Agent Platform Tracing is the production dashboard you use after shipping

Get started with AgentLens

AgentLens is open source and MIT licensed on GitHub. Try it on your own Agentforce agent traces, and if there’s a visualization that would help you debug faster, or a feature you’d like to see added, open an issue.

I built AgentLens because feedback loops matter. Whether you’re debugging Agentforce agents yourself or prompting a coding agent to do it for you, having a clear mental model changes how effectively you work. You write better prompts, catch orchestration issues faster, and gain a deeper understanding of how your agents actually behave.

Resources

• AgentLens web app
• AgentLens VSCode extension
• AgentLens open source repository
• Agent Script Recipes sample app

About the author

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.

The post AgentLens: Debug Agentforce with Interactive Visualizations appeared first on Salesforce Developers Blog.

By leveraging RAG-based semantic retrieval, Abilities solves the growing tool selection problem by identifying the most relevant Salesforce DX MCP Server tools for every request. This approach keeps your execution environment focused while improving output quality and reducing token usage. 

In this post, we’ll examine the challenges of tool explosion, define how Abilities function, and dive into the mechanics of semantic retrieval and tool dependency expansion.

The problem: Tool explosion and manual configuration

In Agentforce Vibes, context is everything. As the Salesforce Platform grows, the number of MCP (Model Context Protocol) servers grows with it, and today, there are more than 80 Salesforce DX MCP Server tools available from within Agentforce Vibes, including task-based tools that handle CLI commands, testing, code analysis, accessibility checks, DevOps operations, mobile, and more.

Turning them all on at once creates problems: the prompt becomes larger, the model has too many choices, output quality drops, and token costs rise. But the opposite problem exists as well. If too few tools are available, Agentforce Vibes may not have the capabilities it needs to generate metadata, create Apex code, refactor LWCs, or successfully analyze code for security violations.

As the catalog of Salesforce DX MCP Server tools expands, several new challenges arise:

• Default limitations: Given the sheer number of possible tools, we’ve only enabled a limited subset by default. This is due to concerns that activating too many tools might degrade the user experience or overwhelm the model.
• User experience impact: This often leads to sub-optimal code generation where the code may fail to compile, fail deployment, or ignore Salesforce development conventions.
• Manual configuration barriers: Developers can customize available tools, but doing so requires manual editing of the local mcp_config.json file and specific knowledge of internal MCP server tool names. Consequently, most developers never make these changes and rely entirely on the default toolset, even when additional tools would improve outcomes.

We needed a way to give Agentforce Vibes exactly the tools it needs, exactly when it needs them.

The solution: Intelligent context activation with Abilities

Abilities are domain capabilities that define what Agentforce Vibes can do in a given development scenario. There is no manual configuration and no switching between development modes. The system responds directly to the developer’s intent. This keeps the execution environment small and focused while ensuring that Agentforce Vibes has the capabilities required to complete the development task.

Examples include capabilities for:

• Apex, LWC development
• DevOps operations
• Mobile development
• Core platform administration

Each Ability bundles together the DX MCP Server tools associated with that capability.

When a request arrives, Agentforce Vibes evaluates the developer’s prompt and determines which capabilities should be active for the task and exposes only the tools associated with those capabilities. This process is managed by an internal orchestration layer called the Ability Coordinator, which interprets the request and prepares the execution environment before the model is invoked. This happens automatically for every task and requires no configuration from the developer.

Abilities are composable and a request may activate more than one capability. In practical terms, Abilities answer one question: What tools should Agentforce Vibes have access to right now?

How Skills and Abilities work together

Abilities and Skills serve different but complementary roles. Abilities handle context activation; they detect what you’re trying to do and automatically select the right Salesforce DX MCP tools for the task. This improves model focus and keeps output aligned with Salesforce best practices.

Skills are reusable, task-level playbooks. They define how specific work gets done (e.g., refactoring a trigger, generating tests, running a deployment), so you don’t restate instructions every time.

Abilities determine what tools are available. Skills define how the work gets done. Together, they let Vibes intelligently configure itself per task while keeping the developer in control.

What drives Abilities

Abilities use retrieval-augmented generation (RAG) to select tools. Instead of sending the full catalog of available tools to the model, Abilities retrieve the most relevant tools using semantic search. This allows the system to scale as the tool catalog grows while keeping prompts small.

How semantic retrieval works

To ensure that Agentforce Vibes selects the right tools without overwhelming the model, we use a RAG-based retrieval process. This happens in three phases: indexing the tools, matching them to your prompt, and then injecting the tools into the model’s execution context.

Phase 1: Tool embedding and indexing

Before any user interaction occurs, we prepare the tool catalog for retrieval. The following process runs whenever the MCP tool catalog is updated:

• Vector conversion: Each tool definition, including its name, description, parameters, and intended usage context, is normalized and converted into a vector embedding.
• Storage: These embeddings are stored in a vector database. This allows the system to identify tools based on their “meaning” and function rather than just matching keywords.

Phase 2: Prompt-to-tool matching (retrieval)

When you submit a prompt, Agentforce Vibes performs a real-time similarity search:

• Prompt embedding: Your prompt is converted into a vector embedding using the same model used for the tool definitions
• Similarity search: The system performs a cosine similarity search, comparing the prompt vector against the vector index of tools to find the best semantic matches
• Top-K selection: To balance performance and accuracy, the system ranks the results and selects the top matches, typically around ten tools

Phase 3: Execution context

These selected tools are then injected into the model’s execution context as a temporary “function catalog.” This ensures the LLM has all the relevant capabilities it needs to fulfill your request while keeping the prompt size optimized for speed and cost.

Handling tool dependencies

Some MCP tools act as orchestrators that guide a workflow and require additional tools to complete the task. If only the orchestrator tool appears in the retrieved set, execution may fail because required tools are missing.

Abilities address this with post-selection expansion.

This ensures that the agent has the tools needed to complete the task while keeping the overall tool set small.

Conclusion

From intent to execution, Abilities ensure that Agentforce Vibes has exactly what it needs, every time. By utilizing RAG-based semantic retrieval and automated dependency expansion, this system effectively manages the expanding catalog of Salesforce DX MCP tools to deliver high-quality, focused development environments without manual configuration.

As Agentforce Vibes grows, the Abilities intelligent context activation system may expand to support:

• Additional Salesforce and third-party MCP servers
• Integration with Skills for task guidance once the right tools are activated
• Richer tool metadata for retrieval
• Improved dependency discovery

Resources

• Agenforce Vibes Developer Guide: Abilities
• Trailhead: Quick Start: Troubleshoot Code with Agentforce Vibes
• Trailhead: Get Started with Agentforce Vibes
• Video: Agentforce Vibes Decoded

About the authors

Jeff Douglas is a Product Management Director at Salesforce working on Agentforce Vibes and AI-powered developer tools. A Salesforce Developer since 2007, and one of the original Salesforce MVPs, he helped launch Trailhead and build several of its core learning and assessment systems. Jeff is an Army veteran, former brewery owner, foster and adoptive parent, woodworker, and owner of a growing herd of mini Scottish Highland cows. Connect with him on LinkedIn.

Ken Lewis is a Lead Member of Technical Services at Salesforce based in Oakland, CA. He formerly worked on Salesforce products for the Nonprofit sector, and enjoys making Salesforce products work best for all types of customers. Outside of work, you can find Ken playing racquetball, practicing on his home DJ setup, or exploring new restaurants and activities in the Bay Area. Connect with him on LinkedIn.

The post Intent-Driven Tool Selection Using Abilities in Agentforce Vibes appeared first on Salesforce Developers Blog.

The feature is based on the OpenTelemetry standard and stores all span data directly in Data 360, making it queryable alongside the rest of your org’s data with a single toggle flip.

How to enable Agent Platform Tracing

Enabling Agent Platform Tracing takes about two minutes:

1. Toggle it on in Setup. Navigate to Setup → Agent Platform Tracing and flip the toggle to enabled. This starts writing span data to Data 360 for every Agentforce action execution in your org.
2. Verify Data 360 is provisioned. Agent Platform Tracing stores spans in data model objects (DMOs), so your org needs Data 360 provisioned and active. If you’re already using Agentforce, this is likely already in place.
3. Check permissions. Users querying trace data need the Data Cloud Data Access permission set, plus read access to the ssot__TelemetryTraceSpan__dlm and related DMOs.
4. Wait for data. Spans begin populating within minutes of the first Agentforce action execution after enablement. There’s no backfill of historical data.

Once enabled, trace data is queryable via SOQL against the Data 360 DMOs described in the sections below.

How Agent Platform Tracing builds a trace tree

The core concept of Agent Platform Tracing is straightforward: each instrumented service that is called when an Agentforce action executes generates a telemetry record called a span, and every span records its parent’s ID. That parent-child chain lets you reconstruct the execution sequence as a hierarchical tree, from the root interaction down through the instrumented downstream calls.

Here’s what that looks like in practice. The following trace shows an agent action executing a flow that results in an error:

[OK] run.interaction (96afcfefaa6fbfe9) — 2,430ms
└── [OK] run.llmstep (9939b8bc33d12bfa) — 167ms
    └── [OK] run.topic.FlowAgentforce__FlowBuilderAutomationsTopic — 3ms
        └── [OK] run.llmstep (b3595456464fbe93) — 1,341ms
            └── [ERROR] run.action.Get_Account_179SB000000v31B — 481ms
                ├── [OK] run.invokeActions.FLOW — 274ms [InvocableAction]
                │   └── [OK] run.Get_Account.1 — 266ms [Flow]
                │       Attributes: flow.api.name=Get_Account, flow.api.version=1,
                │                   flow.execution.api.version=Spring '26,    
   |                    flow.type=AutolnchNoTrig
                │       └── [OK] run.createrecord.account — 42ms [Flow]
                │           Attributes: db.collection.name=Account, db.rows_affected=0,
                │                       db.operation.name=query
                └── [OK] run.llmstep (b4e5194e749ede24) — 884ms

A few key things to call out here: the ssot__DurationNumber__c field (e.g. 2,430ms) tells you how long each step took, and ssot__StatusCode__c (Status is always either OK or ERROR) tells you exactly where a chain broke.

Note that the flow ran successfully — run.Get_Account.1 returned OK — but the parent action still errored. The reason is in the span attributes: db.rows_affected=0 tells you the Account query returned nothing. Despite the operation being named run.createrecord, the db.operation.name=query attribute reveals it was actually performing a lookup, not a write, and that lookup came back empty. The action expected a result and didn’t get one. Execution continued anyway with a fallback run.llmstep, which is why the overall interaction didn’t terminate immediately.

Without the span attributes, you’d see a failed action and have no idea whether the flow ran, whether it touched the database, or what it found there.

Comparing Agentforce session tracing and Agent Platform Tracing

Agent Platform Tracing stores spans into Data Lake Object (DLOs) which get mapped into Data Model Object (DMOs) using the single source of truth (ssot) prefix. Before going further, it’s worth understanding how Agentforce sSession Tracing and Agent Platform Tracing differ.

Agentforce Session Tracing — centered on ssot__AiAgentInteraction__dlm — is instrumented at the planner level and captures the conversational flow: what the user said, which subagent handled the request, and what the agent returned. Agent Platform Tracing, via ssot__TelemetryTraceSpan__dlm, is instrumented at the service level and captures the back-end execution state that powered that response. The span DMO is self-referential by design: ssot__TelemetryParentSpanId__c maps back to ssot__Id__c, enabling tree traversal from any node in either direction. Furthermore, because both DMOs share the exact same underlying Trace ID, you can join ssot__TelemetryTraceId__c directly to ssot__TelemetryTrace__c. This enables you to connect Agentforce session tracing to the specific back-end code execution provided by Agent Platform Tracing.

Querying trace data with SOQL

With SOQL, you can query these DMOs directly via the API to extract exactly the diagnostics you need. A good starting point is profiling performance by operation type, which quickly surfaces any parts of your agent execution that are consistently slow:

SELECT ssot__OperationName__c,
       AVG(ssot__DurationNumber__c) AvgDuration,
       MAX(ssot__DurationNumber__c) MaxDuration,
       COUNT(Id) SpanCount
FROM ssot__TelemetryTraceSpan__dlm
WHERE ssot__StartDateTime__c > 2026-01-01T00:00:00Z
GROUP BY ssot__OperationName__c
LIMIT 20

Or we can query for all recent spans that have errors:

SELECT ssot__Id__c, ssot__OperationName__c, ssot__TelemetryTrace__c,
       ssot__StartDateTime__c, ssot__DurationNumber__c, ssot__StatusCode__c
FROM ssot__TelemetryTraceSpan__dlm
WHERE ssot__StatusCode__c = 'ERROR'
ORDER BY ssot__StartDateTime__c DESC
LIMIT 20

Or count errors across operation types:

SELECT ssot__OperationName__c, COUNT(Id) SpanCount
FROM ssot__TelemetryTraceSpan__dlm
WHERE ssot__StatusCode__c = 'ERROR'
GROUP BY ssot__OperationName__c
LIMIT 20

Once we find a span we want to investigate, we can query for all related spans via ssot__TelemetryTrace__c:

SELECT ssot__Id__c, ssot__OperationName__c, ssot__TelemetryParentSpanId__c,
       ssot__ServiceName__c, ssot__StatusCode__c, ssot__DurationNumber__c,
       ssot__StartDateTime__c, ssot__EndDateTime__c,
       ssot__TelemetrySpanAttributeText__c
FROM ssot__TelemetryTraceSpan__dlm
WHERE ssot__TelemetryTrace__c = 'YOUR_TRACE_ID'
ORDER BY ssot__StartDateTime__c ASC

We can then reconstruct the tree client-side by matching each span’s ssot__TelemetryParentSpanId__c to another span’s ssot__Id__c, treating null or 0000000000000000 as the root.

Query Agent Platform Tracing data from Slack with Slackbot

Because this data lives in Data 360, you can easily query the data to observe Agentforce. One effective pattern is wiring instructions into a Slack canvas, feeding it to Slackbot, and using it as the natural language interface for our trace data. This makes it easy for anyone to digest the trace information, as long as Slackbot is set up to query the information correctly. Whether you’re responding to a user complaint or proactively checking for errors, the workflow is the same: ask the bot, get the trace tree, pinpoint the issue, and discuss a fix, without ever leaving Slack. No log spelunking, no guesswork. Just a conversation with Slackbot to find the root cause and come up with a potential solution.

The approach I take when providing instructions to Slackbot is simple: I provide the schemas, sample queries, join patterns, and main use cases. You can find some of the instructions I use in my  Agent Platform Tracing Slackbot Instructions GitHub repo. 

You’ll need to make sure that your Slack is connected to an org that has Agent Platform Tracing enabled. Once you do, just add the instructions for Slackbot into a canvas and either add the canvas as a tab for Slackbot or paste it into the chat at the beginning of your conversation.

Here’s a typical Slackbot exchange:

You:  Why did interaction b9e8f7b4 take 4 seconds?

Slackbot:  I queried all spans for trace b9e8f7b4. The interaction
           took 2,690ms total. Here's the breakdown:

           run.interaction           2,690ms  OK
           └─ run.llmstep              139ms  OK
              └─ run.topic               3ms  OK
                 └─ run.llmstep        838ms  OK
                    └─ run.action      327ms  OK
                       └─ run.llmstep 1,011ms  OK   ← slowest span

           The bottleneck is the final run.llmstep span (1,011ms),
           which accounts for 38% of total duration.

Once you have Slackbot set up with Agent Platform Tracing, you can ask things like:

• Why did this agent interaction take 4 seconds?
• Which subagent is responsible for the most errors this week?
• What are the most common errors for a given operation type? 
• What’s the root cause of this issue that a user is asking about? 

Joining Agentforce Session Tracing and Agent Platform Tracing data

With the right instructions, Slackbot can query and join both Agentforce Session Tracing and Agent Platform Tracing data together. Joining them provides a full trace tree for end-to-end observability of Agentforce. With both features, we have complete visibility of agent activity and issues in a couple of minutes.

Use this when you want to correlate a user’s conversational context (Agentforce Session Tracing) with the back-end execution details (Agent Platform Tracing). This is an optional advanced use case; both features work independently.

How the join works

Because ssot__TelemetryTraceId__c on ssot__AiAgentInteraction__dlm matches ssot__TelemetryTrace__c on ssot__TelemetryTraceSpan__dlm, they can be used for a direct join. This means you can fetch all spans for an interaction in a single query.

Full joined hierarchy

Session (ssot__AiAgentSession__dlm)
└── Interaction (ssot__AiAgentInteraction__dlm)
    └── Step (ssot__AiAgentInteractionStep__dlm)
        └── Span (ssot__TelemetryTraceSpan__dlm)
            └── Span (ssot__TelemetryTraceSpan__dlm)
                └── ...

Query pattern: Get all spans for a known interaction

SELECT ssot__Id__c, ssot__OperationName__c, ssot__TelemetryParentSpanId__c,
       ssot__StartDateTime__c, ssot__DurationNumber__c, ssot__StatusCode__c
FROM ssot__TelemetryTraceSpan__dlm
WHERE ssot__TelemetryTrace__c = 'TELEMETRY_TRACE_ID_FROM_INTERACTION'
ORDER BY ssot__StartDateTime__c ASC

Getting started

Generative AI adds complexity to modern stacks that traditional logging wasn’t designed to observe, and the cost of those blind spots compounds the longer they take to find. Agent Platform Tracing removes the blind spots. And, when you combine it with Slackbot, it is far easier to observe, troubleshoot, and fix Agentforce issues in production so your agents can get back to work without issues. Check out the Agent Platform Tracing Slackbot Instructions GitHub repo for working Slackbot instructions that you can copy-and-paste into a canvas to get started, and the Agent Platform Tracing documentation for setup instructions and the full data model reference.

Trevor Scott is a Senior Manager of Product Management at Salesforce, focused on enabling customers with security and telemetry data to help developers and admins ship more confidently.

The post Agent Platform Tracing: Debug Agentforce with Trace Trees, SOQL, and Slack appeared first on Salesforce Developers Blog.

This post covers how the new MCP server works, how to install it locally, and how your feedback will help us shape the future of the tool.

Solving the context window challenge

Data 360 has a massive API surface. Exposing every endpoint directly as a tool quickly overwhelms a large language model’s (LLM) context window. To solve this, our engineering team implemented a novel approach inspired by Cloudflare’s research on MCP server tool management at scale. Instead of registering the approximately 200 REST API operations individually, this MCP server consolidates them behind three facade tools:

• search: Discovers tools by intent, keyword, or family.
• payload_examples: Fetches working JSON payloads so the LLM understands how to structure complex requests.
• execute: Runs the specific underlying tool by name and passes the correct parameters.

When an LLM wants to perform a task, it uses a typical workflow: it searches for the right capability, fetches a payload example to understand the required data structure, and then executes the operation. This approach saves context window space and improves the accuracy of the AI’s actions while remaining flexible enough to support the ever-growing headless capabilities of Data 360 far into the future.

What the Data 360 MCP server includes

Under the hood, the three facade tools give your AI client access to hundreds of REST API operations organized into tool families. These tools connect directly to most of the general availability (GA) APIs on Data 360.

Getting started: installation and authentication

During this Developer Preview, the MCP server is designed for local execution and is not yet multitenant. This means each running instance connects a single user to a single Salesforce org.

To run the server, you need Java 17 or later, Maven 3.9 or later, and a Salesforce org with Data 360 enabled. You can use an Agentforce Developer Edition org or a Data 360-enabled Trailhead playground if you want to test it in an isolated environment. The server communicates via stdio, so it will work with any MCP client that supports this transport layer. Once your org is ready, follow the instructions in the README to connect your agent and start building.

The GitHub repo contains an installer script that helps set up dependencies and configure common MCP clients for use with the server.

Help us shape the general availability release

The new Data 360 MCP Server enables you to connect powerful LLMs to your Salesforce data without hitting context window limits. By using a novel facade tool architecture, we’ve made roughly 200 API operations easily accessible for agents and coding assistants. 

The Developer Preview is a critical step before we integrate this architecture directly into the Salesforce platform for general availability. By releasing it as open-source now, we aim to speed up platform integration and gather direct feedback from developers. Install the server today, test it with your local tools, and share your feedback on GitHub to help us prepare for the GA release. 

When the Data 360 MCP server becomes generally available, we’ll offer it as a Salesforce Hosted MCP Server alongside other product integration MCP servers. In the meantime, our entire Data 360 team will be testing it internally alongside you to build and refine a collection of Data 360 agent skills; stay tuned for more news on those!

About the authors

Alba Rivas works as a Principal Developer Advocate at Salesforce. You can follow her on Linkedin.

Chris Peterson is a Senior Director of Product Management at Salesforce, currently working on Data 360 Headless.

The post Introducing the Data 360 MCP Server (Developer Preview) appeared first on Salesforce Developers Blog.

We introduced the newest version of the GraphQL wire adapter (lightning/graphql) in Spring ‘26. Now we’re introducing support for using GraphQL beyond just querying in LWC, adding the ability to easily create, update, and delete records via the GraphQL Mutations API.

Executing GraphQL mutations in LWC with executeMutation

In LWC, interacting with the Salesforce GraphQL API for mutations is achieved via the executeMutation function from the lightning/graphql module.
The executeMutation function allows you to execute one or more data changes. The function accepts an object parameter with the same three properties as the GraphQL wire: query, variables, and operationName. For example, to update an Account’s name:

import { gql, executeMutation } from 'lightning/graphql';

const updateRecord = gql(`
            mutation UpdateAccount($input: AccountUpdateInput!) {
                uiapi {
                    AccountUpdate(input: $input) {
                        success
                    }
                }
            }
`);

 await executeMutation({
            query: updateRecord,
            variables: {
                input: {
                    Id: '001xx000003GYiCAAW',
                    Account: {
                        Name: 'Acme[Updated]',
                    },
                },
            },
        });

The executeMutation function supports dynamic query creation, just like the updated GraphQL support provided by lightning/graphql, so you can build these mutation inputs dynamically with JavaScript.

Data consistency considerations

Lightning Data Service keeps Salesforce record data consistent across UIAPI and GraphQL. With GraphQL mutations, this mostly works seamlessly, but we do have a few tips for getting the best results: 

• Newly created records will not appear in existing GraphQL query results until the query is refreshed, depending on its filter criteria and cache state. This applies to create and update mutations that affect records included in a query result set. 
• Updated record fields may be propagated to subscribed Lightning Data Service wire adapters when their cached data overlaps. It may not be necessary to refresh in all scenarios.
• Deleted records will properly be removed from Lightning Data Service wire results. Refreshes are not required for delete operations.

See it in action: The LWC recipes showcase

To help you get started immediately, we have added dedicated examples to the LWC Recipes GitHub repository.

The new graphqlMutation components demonstrate how to use executeMutation to handle create, update, and delete operations through GraphQL. By exploring the source code for the graphqlMutation components, you can see practical implementations of data mutation, success handling, and error display via toast events.

You can clone the LWC Recipes repository to your local machine and deploy the app to your Salesforce org to test this new functionality.

GraphQL Mutations fully unlock bulk updates and tree saves for Lightning Web Components (LWC) without custom Apex. You can now not only view your Salesforce data efficiently but also modify it seamlessly, bringing richer, data-driven applications to life. This capability significantly streamlines the development of complex UIs that require simultaneous creation, update, or deletion of multiple related records. The integration of GraphQL Mutations marks a major step toward empowering LWC developers with a more modern and powerful way to manage data transactions on the Salesforce platform.

Developer Resources

Ready to start mutating your data? Here are the resources you need to dive in:

• Developer Documentation: lightning/graphql
• GitHub repository: LWC Recipes – Find the new graphqlMutation components and other examples.
• API Developer Guide: GraphQL API – Details on the mutation schema and structure.
• Postman Collection: Salesforce Platform APIs/GraphQL – Helpful for testing mutation structure.
• Best Practices Guide: GraphQL Wire Adapter Best Practices – Maximize your GraphQL potential.

About the Authors

Stephen Carraway is a Lead Member of Technical Staff at Salesforce from the Lightning Data Service team. When he’s not building client data libraries, he’s busy woodworking, gardening, and 3D printing. 

Ben Sklar is a Director of Product Management at Salesforce responsible for UI API, the Salesforce GraphQL API, the Lightning Data Service, and AI Developer Kit. Ben is a major fan of GraphQL, but when not using GraphQL, you can find him playing ultimate frisbee or skiing during the winter. Follow him on LinkedIn.

The post GraphQL Mutations Now Available in LWC: Create, Update, and Delete Records appeared first on Salesforce Developers Blog.

This is where developers get a front-row view of the latest AI innovations, product launches, and breakthrough ideas. You’ll gain the skills to build smarter apps and agents, while connecting with a global community shaping the future of Salesforce development.

Each year, we invite the Salesforce Developer community to take the stage. Whether you’re a first-time Dreamforce speaker or a seasoned pro, this is your opportunity to share your expertise, best practices, and real-world insights with thousands of the highly engaged developers and technical leaders. 

The Dreamforce 2026 Call for Participation is now officially open. If you have a great idea for a session, be sure to follow the guidelines outlined below and complete the submission process by 5 p.m. PT on May 27, 2026.

Why submit a Dreamforce proposal?

Presenting at Dreamforce is about inspiring others, sharing your knowledge and expertise, and helping to build a strong and collaborative community of Salesforce professionals. The developer track is seeking compelling, high-quality session proposals that educate by showcasing expertise through technical deep dives, real-world implementations, and innovative solutions that help attendees learn new skills and drive business outcomes. 

Explore topics for your session

The strongest proposals tell a clear, compelling story: the challenge you faced, the solution you built, and what you learned along the way. To craft a standout submission, consider highlighting one or more of the following:

• Technical deep dives that showcase your expertise – whether it’s a specific feature, a solution you’ve built, or a topic you’re passionate about, grounded in real experience
• The latest Salesforce innovations, including hands-on experience with new or upcoming features
• Pilot or beta experiences that offer early insights into what’s next
• Solutions built with Salesforce technologies such as Agentforce, Data 360, Apex, Lightning Web Components, Model Context Protocol (MCP), vibe coding, and more

Cross-cloud use cases that span multiple Salesforce products – for example, integrations between core Platform and Slack

Choose a session format

For inclusion into the Developer track, you can choose from the following session formats:

• Breakout Sessions (40 minutes): These sessions are held in dedicated rooms with larger audience capacity, giving presenters ample time to deep dive into their topic and take live questions. We recommend 32 minutes for presentation delivery and 5-7 minutes for questions and discussion. Up to three presenters may deliver a breakout.
• Theater Sessions (20 minutes): Fast, focused, and built for impact. Conducted on the show floor in open theaters without formal Q&A, with presenters encouraged to continue the conversation with attendees after their session. Maximum of two presenters.

Submit your proposal 

Submitting a Dreamforce session proposal is a straightforward process. Here are some guidelines for filling out the submission form:

• Review the Call for Participation form before starting. Every detail you include helps the technical review team evaluate your session’s potential. We’re looking for clear storytelling, real-world relevance, and actionable takeaways, so be sure to outline the challenge, your solution, and the technical insights your audience will walk away with.
• Craft an authentic and meaningful title and abstract.
  • Title (60 characters max): Make it very clear to attendees what they’ll learn. Tip: Start with a verb and include the name of the product or technology. Example: “Using Parallel Subscriptions to Help with Your Agents”
  • Abstract (200 characters max): Dive deeper into what your session will cover. Accuracy and clarity are more important than creative writing. Be careful with AI-generated content, which can feel inauthentic and impersonal.
• Make sure you can attend Dreamforce 2026 if your session is accepted. Note that we do not provide travel or accommodation for presenters.

Access speaker support and resources

If your session is accepted, you’ll receive an email confirmation by the end of June 2026, and you and any co-presenters will be asked to confirm your participation via our Content Portal. All presenters receive a complimentary speaker pass to Dreamforce 2026. If you’ve already purchased a pass, you’ll receive instructions on how to request a refund or transfer it to someone else.

We’ll guide you through the session preparation process, which includes coaching you to prepare your content, reviewing your presentation slides, and rehearsing your session. 

Not accepted? Don’t be discouraged. The process is competitive and reflects session balance, audience needs, and event themes, not the quality of your idea. Consider sharing your expertise at Trailblazer Community conferences or Trailblazer Community group meetups.

Important dates to remember

• April 29, 2026: Call for Participation opens
• May 27, 2026, 5 p.m. PT: Submission deadline
• End of June 2026: Proposal status notifications sent
• September 15–17, 2026: Dreamforce 2026, San Francisco

Begin your submission now

Submit your proposal by Wednesday, May 27, 2026 at 5 p.m. PT. We look forward to seeing your ideas.

About the author

Christie Fidura is responsible for Global Developer Programs at Salesforce. You can follow her on BlueSky or LinkedIn.

The post The Dreamforce 2026 Call for Participation Is Now Open appeared first on Salesforce Developers Blog.

This is where the enterprise interface is heading; the AI revolution is here. Today, Salesforce hosted MCP servers are generally available — and they’re the infrastructure that enables that for your organization.

What started as a pilot last spring and beta last October — connecting AI assistants to Salesforce logic and assets through the Model Context Protocol (MCP) — is now a production-ready capability for every Enterprise Edition org and above. Any MCP-compatible client can securely connect to Salesforce with enterprise-grade authentication, governance, and admin control built in.

A hosted MCP server is a Salesforce-managed endpoint that exposes your org’s logic and assets — data, flows, Apex actions, queries, and more — to any AI client that speaks MCP. Salesforce handles hosting, authentication, and permission enforcement automatically. Whether your users live in Slack, Claude, ChatGPT, or something else entirely, MCP means they can work with Salesforce without switching contexts. All the major AI platforms and workplace communication solutions are building on MCP as the standard way to connect AI agents to enterprise systems. We now have an open standard for the industry to converge upon, and Salesforce is enthusiastically embracing the shift.

What’s in the GA

If you read the beta announcement, you saw the vision. For the GA, we’re highlighting the following:

• Fully managed infrastructure. Salesforce hosts and scales your MCP servers, just as it does for the REST APIs you already use. No servers to provision, no uptime to manage. Enable a server in Setup and you’re live.
• Platform security built in. Your existing permissions automatically apply — CRUD, FLS, sharing rules, and all the other controls you’ve already mastered. Every transaction runs as the authenticated user, without anonymous service accounts or a new security model to learn. OAuth and PKCE control access.
• Robust, extensible feature set. Use prebuilt standard servers for the Agentforce 360 Platform, Tableau Next, Data 360 SQL, and more — or configure custom servers that expose your flows, Apex actions, and Named Query APIs.
• Prompt templates. Craft prebuilt starting points for common tasks like account reviews or deal analysis in natural language to guide AI toward useful outputs grounded in your org’s data.

And soon, you’ll be able to open up even more possibilities by invoking Agentforce agents via MCP.

Real-world use cases and impact

We received great customer feedback during the pilot and beta periods regarding use cases for the MCP server:

Sales reps who never leave their AI assistant. A rep preparing for a quarterly business review asks their chatbot to pull the account history, open opportunities, recent case activity, and stakeholder map — all from Salesforce, all in one conversation. No tab-switching, SOQL, or reports to build. The CRM data comes to them.

Cross-system intelligence. A financial services company connects their AI assistant to Salesforce alongside their ERP system. The assistant reconciles quarterly revenue by comparing closed-won opportunities against general ledger entries, all with production data and enterprise-grade security.

Custom tools for domain-specific workflows. ISV partners can build custom MCP tools backed by Apex invocable actions, Apex REST, and more. This gives our mutual customers’ AI assistants the ability to interact with their industry-specific data model through natural language, without the end user needing to understand the underlying schema.

Human at the wheel

In the beta announcement, we characterized MCP as a universal translator between AI agents and your business data. Though that’s accurate, it undersells the importance of enabling secure access to data and actions under human control.

Every MCP transaction runs with the authenticated user’s identity, permissions, and accountability. If the agent updates a record, that person’s name appears in the audit trail. If their permissions don’t allow an operation, the agent can’t do it either. The human is in the driver’s seat.

Steve Jobs called the personal computer “a bicycle for the mind” — a tool that amplified human capability beyond what was possible working by hand. With MCP and today’s AI assistants, we’ve moved past the bicycle. If personal computers were human-powered, now we’ve attached an engine, but the human is still at the wheel. A sales rep preparing for a meeting doesn’t log into Salesforce, navigate to the account, click through related records, and build a briefing by hand. They ask their AI assistant — in whatever tool they’re already working in — and get a comprehensive account review grounded in live customer data.

The pace of business increases every day, and MCP helps your team keep up.

Safety comes faster this time

It took more than 50 years for seat belt mandates in automobiles. In AI, safety controls are evolving in months, not decades.

MCP itself is a safety mechanism. Structured tool calls replace unstructured API access. The server defines exactly which operations are available; there’s no way for an AI assistant to call an API that hasn’t been explicitly exposed. A new OAuth scope provides access to MCP, but not our existing REST APIs. Salesforce adds further layers: the platform’s decades-proven permission model, full audit trails, and the secure-by-default posture requiring MCP servers to be specifically enabled.

Salesforce has always had strong access controls, but that’s not true of every enterprise system. A new category of products is emerging to help manage this on an enterprise-wide basis: MCP gateways — similar to API gateways, but purpose-built for the protocol. Products like MuleSoft AI Gateway provide centralized governance across all your MCP servers, not just Salesforce. Enterprises deploying MCP across multiple vendors will want to explore MCP gateways for centralized access control.

Getting started

The setup takes less than 30 minutes. Here’s the shortest path:

1. Enable a server in Setup → API Catalog → MCP Servers
2. Create an External Client App with mcp_api and refresh_token scopes
3. Connect your client — Postman for testing, then Claude, ChatGPT, or similar for everyday use

Documentation is grouped into three layers to match the needs of various roles:

• Salesforce Help — Admin setup and configuration
• Developer Docs — Technical reference, client setup, and troubleshooting
• Community Wiki — Patterns, tips, and the rationale behind our design decisions

If you’re not sure where to begin, try the platform/sobject-reads server in a sandbox. It’s read-only, risk-free, and immediately useful. When you’re ready for more, you can choose or build the exact tool set your organization needs.

A new chapter

When we launched the beta, we imagined a world where every department could access CRM data through natural language. This year at TDX I presented on how our hosted MCP server brings this to life for Salesforce customers. You can now watch this TDX session on Salesforce+: Master Agentic Integration with Salesforce MCP Servers

CRM is a capability, not a destination. We’re just getting started.

About the author

Ross Belmont is a Senior Director of Product Management focused on integrations, with more than 15 years of experience in the Salesforce ecosystem.

The post Salesforce Hosted MCP Servers Are Now Generally Available appeared first on Salesforce Developers Blog.

Today, we’re announcing two tools that change this. B2C Developer Tooling is an open-source project that|includes a command-line interface (CLI) and a software development kit (SDK) for the entire B2C platform. On-demand sandbox cloning is a new platform capability that lets you create exact replicas of existing sandboxes in minutes. This post covers what the tooling includes, how sandbox cloning works, and how you can get started today.

B2C Developer Tooling: A single CLI for the entire platform

B2C Developer Tooling is an open-source, Apache 2.0–licensed, Salesforce-supported set of packages purpose-built for B2C Commerce developers and admins. We publish three Node.js packages: the CLI (@salesforce/b2c-cli), a reusable SDK (@salesforce/b2c-tooling-sdk), and agent skills for AI-assisted development.

The CLI is built on oclif — the same framework behind the Salesforce sf CLI and Heroku CLI — and provides 200+ commands across 20+ topic areas. That includes code deployment, sandbox management, Managed Runtime, embedded CDN (eCDN) configuration, Shopper Login and API Access Service (SLAS), Custom Salesforce Commerce API (SCAPI) endpoints, Account Manager administration, and more. Whether you’re pushing cartridges with b2c code deploy, tailing Managed Runtime (MRT) logs, or configuring Web Application Firewall (WAF) rules on your eCDN zone — we’ve consolidated it into one tool.

The source code lives on GitHub at the B2C Developer Tooling repository.

Built for the modern platform

We built the tooling SCAPI-first, using Salesforce Commerce APIs wherever available and falling back to the Open Commerce API (OCAPI) only where needed. Authentication is zero-config for most commands — a built-in public OAuth client and browser-based login flow means you can start working without creating API credentials.

For continuous integration (CI) and automation, the CLI is backwards-compatible with sfcc-ci conventions and environment variables, so you can adopt it alongside your existing setup. If you use sfcc-ci, consider updating your CI/continuous delivery (CD) systems to stay up-to-date with the latest platform changes and deprecations. Structured logging with automatic redaction of secrets keeps your pipelines clean and your credentials safe.

Developer velocity

We designed the CLI to keep you moving. Generate new cartridges, controllers, hooks, custom APIs, job steps, and Page Designer components from built-in scaffolds with b2c scaffold generate. Pull instance logs directly from the command line with b2c logs get — and filter by type, level, or time range without navigating WebDAV. For MRT storefronts, b2c mrt tail-logs streams environment logs in real time.

Official GitHub Actions

We provide official GitHub Actions for automating B2C Commerce operations in your CI/CD pipelines. High-level composite actions cover the most common workflows — code deployment, data import, MRT deployment, and job execution — while a raw command passthrough handles everything else. The actions manage CLI installation, credential configuration, and caching automatically. These are fully transparent composite actions defined in YAML, with no compiled JavaScript. See the CI/CD guide for details.

Spotlight: On-Demand Sandbox Cloning

Starting with the 26.3 release, we’re introducing On-Demand Sandbox Cloning — the ability to create exact replicas of existing sandboxes with the same data, configurations, and application code. This is a platform-level capability available through the Control Center UI and the B2C CLI from day one.

Why this matters: environment setup that previously took hours of manual configuration can now be completed in minutes. Need a fresh environment for testing a release? Clone your staging sandbox. Want to reproduce a production issue safely? Clone the relevant environment and investigate in isolation. Each clone is fully independent, so changes in one never affect the other.

Key benefits:

• ~75% faster environment setup, based on internal testing comparing traditional import/export (data import, code deployment, site configuration) to a single clone command 
• Exact replicas — data, site configuration, and code version all carry over
• Fully isolated — each clone is an independent sandbox with its own lifecycle

All of this leads to faster, less error-prone test and development cycles.

When using the cloning feature, you’ll likely still need to update certain environment-specific configurations. Site preferences, aliases, hybrid SLAS configuration settings, and any instances of the tenant identifier of the environment are not automatically updated and should be set with a post-clone data import.

Cloning with the CLI

The CLI provides full support for creating and managing clones:

# Create a clone and wait for it to complete
b2c sandbox clone create zzzv-123 --wait

# List all clones for a sandbox
b2c sandbox clone list zzzv-123

# Get detailed status of a specific clone
b2c sandbox clone get zzzv-123 aaaa-002-1642780893121

You can optionally specify a --target-profile (medium, large, xlarge, xxlarge) and --ttl to control the clone’s resource size and lifetime. The source sandbox is automatically stopped during cloning to ensure data integrity.

Teach your coding agent B2C Commerce

B2C Developer Tooling includes two sets of agent skills designed for AI coding agents:

• b2c-cli — 16 skills covering CLI operations like code deployment, sandbox management, MRT, and eCDN configuration. These teach your coding agent how to use the CLI on your behalf.
• b2c — 18 skills teaching B2C development patterns including controllers, hooks, Custom APIs, Page Designer components, forms, and more. These give your coding agent deep knowledge of B2C-specific architecture and best practices — not just CLI syntax.

The skills work with Claude Code, Cursor, GitHub Copilot, VS Code, Windsurf, and any tool that supports the agent skills format. Install them from the Agent Skills marketplace, through the Skills CLI, or by running b2c setup skills.

The CLI also includes a safety mode that prevents accidental destructive operations, which is particularly important when exposing the CLI as a tool to coding agents. You can set the safety level to NO_DELETE, NO_UPDATE, or READ_ONLY to control what operations are allowed.

Model Context Protocol (MCP) Server

For deeper integration, we also ship an MCP server (@salesforce/b2c-dx-mcp) that gives coding agents direct access to platform tools. It includes tools for managing cartridges, working with MRT environments, and scaffolding Custom SCAPI endpoints. It also supports building storefront-next components with Figma-to-component generation, Page Designer decorators, and site theming. Connect it to any MCP-compatible coding agent for a richer development experience.

Extend with plugins and build with the SDK

The CLI supports an oclif-based plugin system for adding custom commands, and exposes hook points for integrating with external credential stores, modifying API requests, and customizing deployment workflows. Community plugins already exist for IntelliJ configuration, Keychain credential storage, and more.

For custom integrations, the SDK (@salesforce/b2c-tooling-sdk) exposes all core functions as a reusable Node.js library — build your own tools on top of the same APIs that the CLI uses. The CLI also supports shell autocomplete for bash, zsh, fish, and PowerShell.

Conclusion

Between a unified CLI covering the full B2C platform and the ability to clone sandboxes in minutes, we’re making it significantly easier for you to develop, test, and ship on Commerce Cloud. Install the CLI today with npm, try cloning a sandbox, and let us know what you think. Your feedback directly shapes what we build next.

Node.js 22 or later is required. To install agent skills for your coding agent, run b2c setup skills. We welcome feedback and contributions through GitHub Issues.

Resources

• B2C Developer Tooling documentation
• B2C Developer Tooling on GitHub
• Agent Skills guide
• CI/CD with GitHub Actions
• sfcc-ci migration guide

About the author

Charles Lavery is a Principal Member of Technical Staff at Salesforce specializing in developer tooling and APIs.

The post Introducing the B2C Commerce CLI and On-Demand Sandbox Cloning appeared first on Salesforce Developers Blog.