Claude Salesforce MCP server
Register sf-intelligence as a read-only Salesforce MCP server for Claude Code, Claude Desktop, Cursor, Codex, or any MCP client. It is built for org understanding, not metadata deployment.
Last updated 2026-07-29
wire up a client in 2 minutes - no org, no auth
claude mcp add --transport stdio --scope user sf-intelligence-demo -- npx -y sf-intelligence demoRegister this first and your client is talking to a working server before any Salesforce credential is involved - it serves a synthetic "Verdant Energy" org offline, so there is no auth step, no sf CLI, and no DX project to be in. Run /mcp to confirm it connected, ask "what can I ask you?" or "give me an overview of this org", then add the real-org server below once you know the wiring is good. more on the demo →
Claude Code setup
From a Salesforce DX project, register the server once:
claude mcp add --transport stdio --scope project sf-intelligence -- npx -y sf-intelligence mcpThen run a read-only metadata refresh:
npx -y sf-intelligence init
npx -y sf-intelligence refresh --target-org my-org-aliasWorks with MCP clients
| Client | Use case |
|---|---|
| Claude Code | Ask org questions from the project where the metadata vault lives |
| Claude Desktop | Point a desktop assistant at a local read-only Salesforce knowledge base |
| Cursor | Use org metadata intelligence while editing Salesforce code |
| Codex | Give coding agents structured, read-only Salesforce context |
| Any MCP client | Run the same stdio server and reach the same 217 read-only tools — the 19-tool core roster is advertised and directly invokable by default; the rest are reached through sfi.run_analysis |
Why not use a write-capable server for this?
Deployment and retrieval MCP servers are useful for DevOps. sf-intelligence is deliberately different: it is a Salesforce org intelligence layer. It answers questions about dependencies, permissions, automation, Apex, Flows, and documentation from a local vault, and it has no write path to the org.
MCP tools vs Agent Skills
MCP tools and Salesforce Agent Skills are two layers of the same agent stack, not competing choices. The read-only MCP tools are the execution layer: they run grounded, offline reads against your local metadata vault and return cited, confidence-tiered evidence. Agent Skills are the intent layer: portable instructions that teach an assistant when to reach for those tools, how to phrase the question, and how to hedge the answer. Skills interpret what you asked; the MCP layer does the grounded org read.
MCP is not going away here - it stays the grounded execution surface. Skills sit on top of it so the same read-only tools work the same way across Claude Code, Claude Desktop, Cursor, and any other MCP client. See how Agent Skills and MCP tools work together for the full picture.
Searches this page targets
- "Claude Salesforce MCP"
- "Salesforce MCP server Claude Code"
- "Cursor Salesforce MCP"
- "Codex Salesforce MCP"
- "read-only Salesforce MCP server"
- "AI agent Salesforce metadata"
Next steps
For the full setup sequence, read getting started. For how it differs from Salesforce's official DevOps-oriented server, read sf-intelligence vs Salesforce DX MCP Server. For metadata search examples, read Salesforce metadata analysis.
Connect your AI client to Salesforce context.
Register the MCP server, run a read-only refresh, then ask questions from your agent.