A read-only Salesforce MCP server for grounded org intelligence

Last updated 2026-07-23

What is a Salesforce MCP server?

The Model Context Protocol (MCP) lets AI assistants call tools against external systems. A Salesforce MCP server exposes org-aware tools - schema lookup, dependency analysis, permission tracing - that Claude, Cursor, Codex, or any MCP client can invoke during a conversation.

Most Salesforce MCP servers are built for DevOps tasks: deploy, retrieve, run Apex, execute SOQL. sf-intelligence fills the gap they leave open: read-only org intelligence with explicit provenance on every answer.

How sf-intelligence differs

See detailed comparisons on the compare hub — start with vs Salesforce MCP (hosted + DX + Agentforce-as-client).

Try it without a Salesforce org

You do not need an org, an authenticated CLI, or any configuration to evaluate this. One command serves a built-in synthetic org - "Verdant Energy", a fictional solar installer - over MCP. Fully offline, no Salesforce auth, no sf CLI, nothing to connect.

claude mcp add --transport stdio --scope user sf-intelligence-demo -- npx -y sf-intelligence demo

For Claude Desktop, Cursor, Codex, or any other MCP client, the same demo server registers through ordinary stdio config:

{
  "mcpServers": {
    "sf-intelligence-demo": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "sf-intelligence", "demo"]
    }
  }
}

Then ask it something a real org question would look like - "what breaks if I delete Widget__c.Status__c?" - and read the citations it returns. When you want it pointed at your own org, the requirements below apply.

Works with Claude, Cursor, Codex, and more

Register once via stdio transport. The server resolves the vault from your Salesforce DX repo's working directory. For a step-by-step walkthrough, see Claude Salesforce MCP setup.

Claude Code (project-scoped)

claude mcp add --transport stdio --scope project sf-intelligence -- npx -y sf-intelligence mcp

Claude Desktop / Cursor / any MCP client

{
  "mcpServers": {
    "sf-intelligence": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "sf-intelligence", "mcp"]
    }
  }
}

Remote read-only HTTP (optional)

For CI or shared read-only access on localhost, sfi serve --http exposes streamable HTTP with bearer-token auth. The live plane is hard-disabled over HTTP - remote callers cannot reach your org. Details in the configuration reference.

217 read-only tools, plain-language routing

You never memorize tool names. Ask questions like:

• "What breaks if I delete the Contact email field?" → impact analysis
• "Why can't this user see this record?" → sharing troubleshooting
• "Where is this Apex class used?" → dependency graph + source grep
• "Generate a data dictionary for this org" → documentation tools

An offline semantic router surfaces a ranked shortlist of the right tools for each question across offline, live, and hybrid planes, and your AI host picks which to run. It fails closed: write requests and prompt injection are refused with a read-only alternative, and a question nothing covers gets an honest gap instead of a lookalike tool. Browse the full capability map or tool reference.

How MCP tools and Salesforce Agent Skills complement one another

MCP tools and Salesforce Agent Skills solve different halves of the same problem, and neither replaces the other. Skills capture intent and workflow - what an admin, developer, or architect is trying to accomplish, and the order to do it in. The read-only MCP tools do the grounded execution - resolving a component, tracing dependencies, and returning evidence-labelled answers from your local vault. Skills call the MCP layer; the MCP layer keeps every step grounded in real metadata. MCP stays the read layer that answers the questions; skills make those answers repeatable.

Safe for production orgs

Enterprise testers' first question: what does this do to my org? The answer is nothing destructive:

• Only sf project retrieve touches Salesforce - metadata read, never write
• Vault answers make no network calls
• Live plane is off until you grant per-org consent
• No telemetry, no phone-home

Full guarantees on the trust page and in FAQ.