Sf Agentforce Mcp Actions

Agentforce MCP Server Actions

How to use Model Context Protocol (MCP) servers as Agentforce action types. For building MCP servers, see mcp-server-patterns.

When to Use

• Registering an MCP server for use in Agentforce topics
• Configuring auth, rate limits, or tool discovery for MCP actions
• Using prebuilt Salesforce MCP servers (DX, Heroku, MuleSoft)
• Setting up Hosted MCP (Pilot) for zero-infrastructure endpoints
• Debugging MCP action resolution in Agent Script

---

Overview

MCP exposes external tools to Agentforce via JSON-RPC 2.0 over HTTP/SSE. An MCP server declares tools (with names, descriptions, parameters, return types), and Agentforce discovers them at connection time. Each tool becomes an available agent action.

---

Setup

1. Register in MCP Server Registry

Setup > MCP Servers > New MCP Server

| Field | Value | |---|---| | Name | Descriptive name (e.g., "Weather API") | | Endpoint URL | Server URL (HTTPS required for production) | | Auth | OAuth 2.0 with Integration User | | Transport | HTTP/SSE (standard) |

After registration, tools appear in Agentforce Asset Library and can be added to topics.

2. Add to Agent Script

In Agent Script, MCP tools are referenced like any other action. The target: field is not needed — MCP tools are auto-discovered from the server's tool manifest.

topic weather:
   actions:
      get_weather:
         description: "Get current weather for a city"
         inputs:
            city: string
               description: "City name"
               is_required: True
         outputs:
            temperature: number
               description: "Temperature in Fahrenheit"
               is_displayable: True

   reasoning:
      actions:
         weather: @actions.get_weather
            with city = ...

3. Classic Setup (Agent Builder UI)

Setup > Agentforce > Agent Assets > Add Action > select MCP tool from Asset Library.

---

Auth

• OAuth 2.0 with Integration User (least privilege)
• FLS and sharing rules enforced on the Integration User
• External credentials managed via Named Credentials
• Token refresh handled automatically by the platform

---

Rate Limits

| Limit | Value | |---|---| | Requests per minute per server | ~50 | | Timeout per tool call | 120 seconds (matches agent timeout) | | Max payload size | Platform-dependent |

---

Tool Discovery

On connection, the MCP server returns its tool manifest:

{
  "tools": [
    {
      "name": "get_weather",
      "description": "Get current weather conditions",
      "inputSchema": {
        "type": "object",
        "properties": {
          "city": { "type": "string", "description": "City name" }
        },
        "required": ["city"]
      }
    }
  ]
}

Agentforce uses tool names and descriptions for LLM routing — keep them clear and specific.

---

Prebuilt MCP Servers

| Server | Purpose | Setup | |---|---|---| | Salesforce DX MCP Server | Deploy, test, manage scratch orgs from AI assistants | @salesforce/mcp npm package | | Heroku Platform | Manage Heroku apps, dynos, add-ons | Built-in connector | | MuleSoft | API orchestration, integration flows | MuleSoft Anypoint connector |

---

Hosted MCP (Pilot)

Fully managed cloud endpoints — zero infrastructure. Pre-built for core CRM and B2C Commerce APIs.

• No server to deploy — Salesforce hosts the MCP endpoint
• Admin setup only — enable in Setup, configure permissions
• Pre-built tools for standard CRM operations (account lookup, case creation, etc.)
• B2C Commerce tools for product catalog, order management

---

Debugging

If an MCP action fails in Agent Script:

1. Check server connectivity: Verify endpoint URL is reachable from Salesforce
2. Check auth: Ensure Named Credential + External Credential are configured
3. Check tool name: The action name in Agent Script must match the tool name in the server manifest
4. Check rate limits: Monitor for 429 responses in MCP Server logs
5. Test standalone: Use sf mcp test or direct HTTP to verify the server responds

---

Related

• Skill: mcp-server-patterns — building MCP servers (Node SDK)
• Skill: sf-agentforce-development — Agent Script patterns
• Reference: @../reference/AGENTFORCEPATTERNS.md

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jiten-singh-shahi
• Source: jiten-singh-shahi/salesforce-claude-code
• License: MIT
• Homepage: https://www.npmjs.com/package/scc-universal

Install and usage instructions live in the source repository linked above.