- Home
- Salesforce integration
- Salesforce MCP
Salesforce MCP: connect Claude, ChatGPT and other AI assistants to your Salesforce data
MCP (Model Context Protocol) is an open standard. With it, an AI assistant like Claude or ChatGPT can use the tools in another app. A Salesforce MCP server is how it reads and changes your data in Salesforce. This page shows which servers there are, who makes them, what ended on 2 November 2026 and how to set one up with care.
The short answer
Salesforce makes more than one MCP server, and each is at its own stage. Its hosted MCP servers run inside Salesforce, not on your computer. Its GitHub page says they are now generally available (GA), which means a full release. Its local DX server, a tool for developers, ended on 2 November 2026. Salesforce in Claude was still for pilot customers when we read its page on 18 November 2026. Those are a group Salesforce chose to try it first.
Salesforce's GitHub page says its hosted MCP servers are now generally available. Each one is turned on in Setup, Salesforce's admin settings.
Source: Salesforce on GitHub, Salesforce Hosted MCP Servers (forcedotcom/mcp-hosted). Read 19 Nov 2026The day Salesforce's local DX MCP Server reached end of life. Its two GitHub repositories are now read-only.
Source: Salesforce on GitHub, DX MCP Server end of life notice. Read 18 Nov 2026Ready-made sales skills in Salesforce in Claude, paired with a prebuilt Salesforce MCP server. For pilot customers when we read its page.
Source: Salesforce, AI and Salesforce in Claude. Read 18 Nov 2026Salesforce’s MCP servers, and who makes each one
Salesforce offers several MCP routes, and each one does a different job. Each card names the maker, and the status where the route’s own page gives one, as we read it between 15 and 19 November 2026. The last card covers servers Salesforce does not make.
- Hosted MCP servers, from Salesforce. Salesforce’s GitHub page for them says they let AI assistants like Claude and ChatGPT connect securely to your Salesforce logic and assets. It says they are now generally available (GA). Salesforce’s overview adds that they are intended only for customers with Flex Credits, the units in its pay-per-use pricing for Agentforce, and that you may be billed for server usage. Each server is turned on in Setup. Some servers’ tools are still in beta, as the setup section below shows.
- Salesforce in Claude, from Salesforce and Anthropic. Salesforce says it pairs a prebuilt Salesforce MCP server with 37 ready-to-use sales skills. A skill is a packaged set of steps and scripts an AI tool can follow. It comes from Claudeforce, a partnership between Salesforce and Anthropic. Its page still said pilot customers on 18 November 2026.
- Agentforce Coworker, inside Claude and ChatGPT. Agentforce is Salesforce’s own AI agent platform. Salesforce calls Agentforce Coworker “your autonomous AI teammate”, and says you can use it in Claude and ChatGPT through MCP, as well as in Salesforce, Microsoft Teams and on mobile.
- Headless Toolkit, for developers. Salesforce says the release has over 60 new MCP tools and 30+ coding skills. It connects coding tools such as Cursor and Claude Code to Salesforce.
- Tableau MCP, for Tableau Cloud. Tableau’s own server runs as a managed service at mcp.tableau.com. Each user signs in with their own Tableau Cloud login through OAuth 2.1, a standard way to sign in without giving the app your password. Our page on Salesforce reports and dashboards covers the reporting side.
- MuleSoft MCP Connector 1.7, for your own APIs. It turns your APIs (the doors your systems open so other software can ask them for data), connectors and apps into tools an AI client can call. A Mule app can act as an MCP server or an MCP client, and it runs inside Mule, so there is no hosting of your own to set up.
- The DX MCP Server, now ended. Salesforce’s local server for developers reached end of life on 2 November 2026. The section after next covers what changed.
- Community servers, not made by Salesforce. One example is tsmztech/mcp-server-salesforce on GitHub. Its own page calls it an MCP server that connects Claude with Salesforce data and metadata.
We read Salesforce’s overview of its hosted servers on 15 November 2026, but could not load the rest of its developer site, so the full list of hosted servers and their tools is not on this page. The setup section below covers one server.
One request through Salesforce in Claude, step by step
Here is one request moving through Salesforce in Claude. The deal and the request are made up. The rules come from Salesforce’s page, plus two from the MCP standard (steps 3 and 7).
- An admin connects Salesforce in Claude once, and it works for the whole team.
- A sales rep asks Claude to update the next step on one of her deals and to email the buyer.
- Claude asks the Salesforce MCP server which tools it offers. MCP requires a server with tools to answer that request.
- Claude reads the deal. It sees only what the rep may see in Salesforce, “no more”, in Salesforce’s words.
- Claude updates the next step. Salesforce says it changes only the field it said it would.
- The buyer works at another company. Claude can check with the rep before it emails anyone outside your company, or just send it once she is comfortable with that.
- The MCP standard adds that a person should always be able to deny a tool call.
Claude never had more access than the rep. Check one thing before you plan around it. On 18 November 2026, Salesforce’s page still said Salesforce in Claude was “available today for pilot customers” and “planned for open beta in September 2026”. We found no newer date on the pages we read. The same page labels its sales skills Beta and offers a beta sign-up.
What ended on 2 November 2026: the local DX MCP server
On 2 November 2026, Salesforce’s local DX MCP Server reached end of life. Salesforce’s notice on GitHub says that after that date it gets no fixes, no support and no new releases.
The DX server ran on a developer’s own computer. It worked only with orgs that were authorized on that computer first. An org is one company’s Salesforce account. The server’s README, the project’s read-me file, says it includes over 60 tools. One of them, run_soql_query, runs a SOQL query against an org. SOQL is Salesforce’s query language.
- The code is frozen. The code repository, salesforcecli/mcp, and its issues repository, forcedotcom/mcp, were both archived on 2 November 2026 and are now read-only.
- Old setups still start. Salesforce’s notice said the install package on npm would be marked deprecated but stay installable, so setups pinned to version 0.30.15 keep working. They get no further fixes.
- Skills take its place. Salesforce says its tooling “is now delivered as skills”. A skill packs steps, scripts and reference notes into one folder that an AI tool can use. Salesforce’s are in a repository called forcedotcom/sf-skills, with plugins and sample apps.
- Skills can change between releases. Salesforce warns that skills and plugins may be renamed, restructured or removed between releases. They do not carry the stability promise of its generally available APIs.
- The hosted servers are separate. The notice names the local server only. Salesforce’s GitHub page for its hosted servers says they are generally available.
If someone on your team still runs the DX server, check which orgs it can reach. Its README says the ALLOW_ALL_ORGS value opens every authorized org, and to use it with caution. Old advice lingers, too. Salesforce’s Agentforce developers page still named the DX server on 19 November 2026.
Set up one hosted server: the API Context server
Salesforce’s Metadata API guide walks through one hosted server, the API Context server, from Setup to a first test. Salesforce calls it one of many hosted MCP servers.
It is a tool for people who build on Salesforce. Its tools tell an AI tool how each kind of Salesforce metadata is built. They give complete field definitions, valid values, constraints and examples, so the AI writes accurate metadata files. Metadata is the set of files that describes how an org is set up, such as its custom objects and fields. Salesforce lists the tools for Professional, Enterprise, Unlimited and Developer orgs, sandboxes and scratch orgs that have API access turned on. The guide writes its app steps for Cursor. Only step 7, the Claude part, comes from its section Claude Desktop in Developer Mode.
- In Setup, type MCP Servers in the Quick Find box and open MCP Servers. Click salesforce-api-context, then click Activate.
- In External Client App Manager, create a new external client app, the record that lets an outside app connect to Salesforce. Salesforce says you need one for any MCP client other than Agentforce Vibes, its developer tool that runs inside VS Code.
- Turn on OAuth for the app. Enter the callback address your AI client’s maker gives. That is the web address the sign-in returns to. Add the scopes Salesforce lists, which are the permissions the app asks for: api, sfap_api, and refresh_token with offline_access. If you use prompt templates, add a fourth, einstein_gpt_api.
- Under Security, pick JWT-based access tokens for named users. These are sign-in passes in a standard format. Also require PKCE, an extra check during sign-in, and turn off every other option.
- Copy the app’s consumer key, its ID for sign-in.
- Log out of every other Salesforce org. Then, in your default browser, log in to the org where you made the app, and leave the browser open. The AI client opens a new tab in it later. Salesforce recommends these steps because the MCP spec has no special rules for systems with many tenants, like Salesforce.
- In Claude Desktop, install Salesforce’s extension file for hosted MCP servers. Enter the server name platform/salesforce-api-context, paste the consumer key, then turn the extension on with its toggle.
- Test the link with a simple prompt, such as asking for the metadata context of one metadata type.
These are the steps for this one server. Other hosted servers can need other settings, so check Salesforce’s page for the server you pick. The guide gives this server’s address in two forms:
- For a production org:
https://api.salesforce.com/platform/mcp/v1-beta.2/platform/salesforce-api-context - For a scratch or sandbox org:
https://api.salesforce.com/platform/mcp/v1-beta.2/sandbox/platform/salesforce-api-context
Salesforce’s guide names two limits for this server. Its Metadata API Context tools are a beta service. And the API calls this feature may make to your org count against your org’s API quota, the number of API calls your org may make in a day.
Salesforce’s Event Monitoring logs can also mark which API requests came from a hosted server. To read them, a user needs the View Event Log Files and API Enabled permissions. Their API Total Usage event covers REST, SOAP and Bulk API requests. In it, the client category SALESFORCE_HOSTED_MCP marks a request from a hosted MCP server.
Tell us which AI assistant should work in your Salesforce org
The first call is free. Bring the assistant your team uses and the Salesforce jobs you want it to do, and we can talk through which MCP server fits.
Which Salesforce MCP route fits your job
Start with the job, then pick the narrowest route that does it. This table is our view, built from the vendor pages listed on this page.
Scroll the table sideways to see every column.
| If you want | Start with | Why |
|---|---|---|
| An assistant your team picks, such as Claude or ChatGPT, working with Salesforce data | Salesforce’s hosted MCP servers, if your org has Flex Credits | Generally available, but intended only for customers with Flex Credits, in Salesforce’s words; each server is turned on in Setup |
| Sales reps working deals inside Claude | Salesforce in Claude, once it leaves pilot | A prebuilt server with 37 sales skills, connected once by an admin |
| Developers building from Cursor or Claude Code | Headless Toolkit and the sf-skills repository | Over 60 MCP tools, and skills replace the DX server |
| Analysts asking questions of Tableau Cloud | Tableau MCP | Each user signs in with their own login |
| Agentforce agents that use another company’s tools | A server from AgentExchange, deployed in Agent Builder | No code, in Salesforce’s words |
| An assistant that also needs your ERP or other APIs | MuleSoft MCP Connector, or a custom MCP server | Your APIs become tools an AI client can call |
| Rules only your company has | A custom MCP server or agent | It calls Salesforce’s REST API under your own rules |
Seven rules for a safe Salesforce MCP setup
Each rule comes from Salesforce’s pages, Tableau’s page or the MCP specification. Each card says what to do and why.
- Give the assistant only the rights it needs. Salesforce in Claude sees and does only what the signed-in user may. Tableau MCP applies each user’s own permissions. A Headless Toolkit agent runs under its own identity, which follows your existing permissions, field-level security and sharing rules, Salesforce’s rules for which fields and records each person may see.
- Turn on only the tools the job needs. Salesforce’s README for the DX server, before it ended, warned that turning on all of its 60-plus tools can overwhelm the AI model’s context, its working memory for a task. By default, that server used only tools marked generally available.
- Sign in through an external client app. Salesforce requires an external client app or a connected app, with an OAuth 2.0 flow, for any outside tool that reaches its REST API. Creating connected apps is restricted as of Spring ’26, and Salesforce recommends external client apps.
- Keep a person in the loop. The MCP standard says a human should always be able to deny a tool call. Salesforce in Claude can ask before it emails anyone outside your company.
- Count the API calls. Salesforce counts REST, SOAP and Bulk API calls toward one daily allocation for the whole org, not per user. API use by the hosted API Context tools counts against your org’s API quota too.
- Do not trust a tool’s own description. The MCP specification says descriptions of a tool’s behavior, such as annotations (labels a tool gives itself), should be treated as untrusted unless they come from a trusted server.
- Check how a community server signs in. tsmztech’s server offers three sign-in methods, and username and password is the default. Its OAuth option starts by creating a connected app, a kind of app whose creation Salesforce restricts as of Spring ’26. It recommends Salesforce CLI login for local and development use.
Is MCP just an API? No, it sits on top of one
MCP is a standard for how an AI app finds and calls tools. Behind each server, something still has to read and write the Salesforce records.
In MCP’s terms, the host is the AI app you use, the client is the connector inside it, and the server offers tools and data. Each client talks to exactly one server. A server should not be able to read the whole conversation or see into other servers. The MCP docs call MCP “an open-source standard for connecting AI applications to external systems”. Salesforce’s page calls it “USB-C for AI”.
A server can offer three things. Tools are actions the model can run. Resources are data it can read. Prompts are ready-made templates. The messages travel as JSON-RPC 2.0, a plain format for one program to ask another to do something.
The table below is for developers. It shows the two ways a client and a server talk, and how each one signs in.
Scroll the table sideways to see every column.
| Transport | How it works | How it signs in |
|---|---|---|
| stdio | Your client starts the server on your own computer | The spec says it should read credentials from the environment. The DX server used orgs authorized on that computer |
| Streamable HTTP | Each message is an HTTP POST to one web address | A protected server acts as an OAuth 2.1 resource server, which checks each sign-in pass. Tableau’s hosted server signs users in with OAuth 2.1 |
MuleSoft’s connector supports Streamable HTTP, which MuleSoft recommends, and SSE, an older way to stream replies, which it marks deprecated.
MCP runs both ways in Salesforce
Salesforce is the server when Claude or ChatGPT reaches into your org. It is the client when an Agentforce agent reaches out to another company’s server.
Salesforce’s MCP page says you pick a server from AgentExchange, a curated catalog of vetted services. You deploy it in Agent Builder, Salesforce’s tool for building agents, with no code. The same page says Agentforce will deliver an AI agent gateway. It would keep one central list of all MCP servers and limit the rate of calls to prevent overload. The page gives no date, so check what your own org has.
When a custom MCP server or agent fits
Use Salesforce’s own route when it covers your job. A custom server fits when the job crosses into another system, such as your ERP, or follows rules only your company has.
A custom server would call Salesforce’s REST API. So it needs an external client app, or a connected app you already have, with an OAuth 2.0 flow. Its calls count toward your org’s daily allocation. If a paid org goes over, Salesforce lets calls run past the limit by a certain amount, if possible, up to a hard cap. Trial orgs, Developer Edition orgs and sandboxes get no such room.
Our team has built Salesforce agents in Agentforce and as custom agents through Salesforce’s API. Separately, our team takes Salesforce implementation end to end: the org set up, the data moved in and the team trained.
This page covers the link between an AI assistant and Salesforce. For the jobs an agent does inside Salesforce, see Salesforce AI agents. For MCP across many vendors, see what an MCP server is. For links between Salesforce and your ERP or other apps, see Salesforce integration. For the same question in other systems, see HubSpot MCP and QuickBooks MCP.
The first call is free. We sign your NDA before you share any data, if you ask.
Questions about Salesforce MCP
What is Salesforce MCP?
It means linking an AI assistant to Salesforce through MCP, an open standard first developed by Anthropic. Salesforce offers several routes. Its hosted MCP servers are generally available. Salesforce in Claude was for pilot customers when we read its page. The local DX MCP Server ended on 2 November 2026.
What is MCP access?
It is the access an AI assistant gets to an app through an MCP server. In Salesforce, it is also an OAuth scope, which is a named permission a sign-in can ask for. Salesforce's Metadata API guide lists a scope value called MCP that allows access to Model Context Protocol. In Salesforce in Claude, the assistant sees only what the signed-in user may. Tableau MCP uses each user's own permissions. A Headless Toolkit agent gets its own identity instead, which follows your Salesforce permissions.
How to connect to Salesforce MCP?
It depends on the server. In Salesforce's quick start for one hosted server, the API Context server, you turn the server on in Setup, create an external client app with OAuth turned on, and give its consumer key to the AI client. Other hosted servers can need other settings. For Salesforce in Claude, an admin connects it once for the whole team. For Tableau's hosted server, you point your MCP client at mcp.tableau.com and sign in with your own Tableau Cloud login.
Are Salesforce's hosted MCP servers free?
Salesforce's own overview says use may be billed. As we saw it on 15 November 2026, it said the hosted servers are intended only for customers with Flex Credits, and that you may be billed for server usage. Flex Credits are the units in Salesforce's pay-per-use pricing for Agentforce. Calls that one hosted server makes to your org also count against your org's API quota. Ask your Salesforce account team what your contract covers.
Is MCP just an API?
No. MCP is a standard way for an AI app to find and call tools, and it sits on top of a system's own API. Behind the server, something still has to read and write Salesforce records. A custom server would call Salesforce's REST API, and those calls count toward one daily allocation for the whole org.
Is the Salesforce DX MCP Server still supported?
No. Salesforce says the local DX MCP Server reached end of life on 2 November 2026. It gets no more fixes, support or releases, and its GitHub repositories are now read-only. The npm package still installs, and setups pinned to version 0.30.15 keep working. Salesforce says its tooling now ships as skills, in the forcedotcom/sf-skills repository. Its hosted MCP servers are a separate product.
Can Claude or ChatGPT work with Salesforce through MCP?
Yes. Salesforce's GitHub page says its hosted MCP servers let assistants like Claude and ChatGPT connect to your Salesforce logic and assets. Agentforce Coworker also works in Claude and ChatGPT through MCP, Salesforce says. Salesforce in Claude was still for pilot customers on 18 November 2026.
Does Agentforce support MCP?
Yes, both ways. Agentforce can call other companies' MCP servers. Salesforce says you pick one from AgentExchange, a curated catalog of vetted services, and deploy it in Agent Builder, its tool for building agents, with no code. Agentforce Coworker also works inside Claude and ChatGPT through MCP, Salesforce says. Agent Builder can link agents to Flows and Apex (Salesforce automations and code), MuleSoft APIs and prompt templates too.
Is there a Salesforce MCP server on GitHub?
Yes, a few. Salesforce keeps a GitHub page for its hosted MCP servers, at forcedotcom/mcp-hosted, and it says they are generally available. The DX MCP Server's two repositories were archived on 2 November 2026 and are read-only. Salesforce's skills live in forcedotcom/sf-skills. Tableau's server is in tableau/tableau-mcp. Other Salesforce MCP servers on GitHub, such as tsmztech/mcp-server-salesforce, come from the community.
Is Salesforce on AWS or Azure?
On AWS, Google Cloud or Salesforce's own servers. Salesforce's sub-processors document says a Sales Cloud or Service Cloud org runs on one of the three, in a region picked when the org is created. The same document lists a service called the MCP Server among features processed on AWS. It names Microsoft Azure too, as a provider of generative AI services.
Who can set up a Salesforce MCP server for us?
A team that knows Salesforce's sign-in rules and its API. Our team has built Salesforce agents in Agentforce and as custom agents through Salesforce's API. Separately, it takes Salesforce implementation end to end: the org set up, the data moved in and the team trained. The first call is free. We sign your NDA before you share any data, if you ask.
Sources
29 pages, read 15 to 19 Nov 2026: every source behind a figure or a vendor fact on this page.
- Salesforce on GitHub, Salesforce Hosted MCP Servers (forcedotcom/mcp-hosted)Read 19 Nov 2026
- Salesforce Developers, Hosted MCP Servers overviewRead 15 Nov 2026
- Salesforce, Metadata API Developer Guide (PDF)Read 18 Nov 2026
- Salesforce, Object Reference (PDF)Read 18 Nov 2026
- Salesforce on GitHub, DX MCP Server end of life noticeRead 18 Nov 2026
- Salesforce on GitHub, forcedotcom/mcpRead 18 Nov 2026
- Salesforce on GitHub, DX MCP Server READMERead 18 Nov 2026
- Salesforce on GitHub, sf-skillsRead 18 Nov 2026
- Salesforce, Agentforce for developersRead 19 Nov 2026
- Salesforce, ClaudeforceRead 18 Nov 2026
- Salesforce, AI and Salesforce in ClaudeRead 18 Nov 2026
- Salesforce, Agentforce CoworkerRead 18 Nov 2026
- Salesforce, Agentforce pricingRead 18 Nov 2026
- Salesforce, Headless ToolkitRead 18 Nov 2026
- Tableau on GitHub, Tableau MCPRead 18 Nov 2026
- MuleSoft docs, MCP Connector 1.7Read 18 Nov 2026
- Salesforce, Agentforce MCP supportRead 18 Nov 2026
- Salesforce, Agentforce BuilderRead 18 Nov 2026
- Salesforce, Platform limits cheat sheet (PDF)Read 18 Nov 2026
- Salesforce, REST API Developer Guide (PDF)Read 18 Nov 2026
- Salesforce, Infrastructure and Sub-processors (PDF)Read 18 Nov 2026
- Model Context Protocol, introductionRead 18 Nov 2026
- Model Context Protocol, specificationRead 18 Nov 2026
- Model Context Protocol, architectureRead 18 Nov 2026
- Model Context Protocol, transportsRead 18 Nov 2026
- Model Context Protocol, authorizationRead 18 Nov 2026
- Model Context Protocol, toolsRead 18 Nov 2026
- tsmztech on GitHub, mcp-server-salesforce (community, not Salesforce)Read 18 Nov 2026
- CadenceWorks, Salesforce AI agentsRead 18 Nov 2026
Tell us the job you are weighing
Name the system you run and the work you have in mind. A person on our team replies by email.
Illustration