Two Salesforce MCP servers, and only one is current
Searching for “Salesforce MCP server” turns up two different things:
| Hosted MCP servers | Salesforce DX MCP server | |
|---|---|---|
| For whom | Anyone working with Salesforce data and logic through an AI tool | Developers working on Salesforce code and metadata |
| Where it runs | On Salesforce, at https://api.salesforce.com/platform/mcp/v1/... |
Locally, npm package @salesforce/mcp |
| Status | Generally available | End of life since November 2, 2026; replaced by skills in forcedotcom/sf-skills |
This page is about the hosted servers. If you still run the DX server, it keeps working at the pinned version, but it gets no further releases.
Which hosted server to start with
Salesforce’s own FAQ gives a clear starting point:
platform/sobject-reads: read-only. No risk of changing data. Good for a first test, a sandbox and building trust.platform/sobject-all: read, write and delete, plus prompt templates (usable in Claude and Cursor) for tasks such as an account review briefing.- A custom server: you bundle only the tools you want, including your own APIs, Flows and Apex.
Setup, step by step
- Check the edition. Enterprise Edition or above, or a Developer Edition org.
- Activate the server in Setup, API Catalog, MCP Servers. At GA every server is off by default and must be switched on one by one. A change can take up to two minutes.
- Create an External Client App (not a Connected App) with scopes
mcp_apiandrefresh_token, PKCE on. It can take up to 30 minutes before it works everywhere. - Connect the client. Claude (web), ChatGPT and Postman need no local install; Cursor and Claude Desktop do. Use the production URL or the separate sandbox URL (
/v1/sandbox/...). Themcp-remotebridge library is not supported. - Test with the
getUserInfotool. A 404 usually means the server is not activated; a 401 means wrong scopes or an External Client App that has not propagated yet.
The common misunderstanding: “the AI gets access to Salesforce”
It gets access as a person. The FAQ is precise: the agent gets “the same permissions as the human who authenticates”, no more and no less, including object access, field-level security and sharing rules. If an agent updates a record, the audit trail shows the user’s name as editor.
Two consequences. First, a sales rep’s agent cannot see what the rep cannot see, which is good. Second, an admin’s agent can do everything the admin can do, which is why you start with sobject-reads and narrow access with a custom server or with profile and permission set assignments on the External Client App. Also keep in mind that tool calls count against your daily API quota.
What this means for a team
With hosted MCP, each person’s Claude or ChatGPT can look up accounts, opportunities and cases in Salesforce. But Salesforce holds records, not the reasoning around them: why a deal was discounted, what a customer hinted at on a call, how your team qualifies a lead. That knowledge stays in heads and chat threads, and each person’s AI starts from zero. See how to connect MCP to Claude, ChatGPT, Copilot and Grok for client setup, and Salesforce Agentforce or Bombos for how an agent platform inside Salesforce compares with a shared memory layer.
Where Bombos fits
The hosted MCP server gives every AI tool access to your Salesforce records; Bombos gives them the same memory of how your team sells. It is a plugin underneath Claude, ChatGPT, Copilot, Grok and your own agents that learns how your team works, draws on what already sits in mail and documents, and shows with every answer what it rests on. It runs in the cloud in Europe or standalone on your own server, and one person can try it first with their own agents.
Related questions
Which Salesforce editions can use the hosted MCP servers?
Enterprise Edition or above, plus Developer Edition orgs. If you have the right edition but the service is missing, Salesforce Support can enable the MCPService permission for your org.
Can I use a Connected App?
No. Hosted MCP only works with External Client Apps, with the scopes mcp_api and refresh_token.
Do MCP calls count against our API limits?
Yes. Each tool call uses one or more API calls from your org's daily quota. Check Setup, Company Information, API Requests, Last 24 Hours.
What happened to the Salesforce DX MCP server?
The local developer server @salesforce/mcp reached end of life on November 2, 2026. It still installs, but gets no updates; Salesforce now ships that developer tooling as skills in forcedotcom/sf-skills.
Does Salesforce run an LLM behind the MCP server?
No. According to the project FAQ the server is fully deterministic: it takes structured tool calls and returns data. The model runs entirely in your client.
