Agentforce MCP support means Salesforce's AI agents can use the Model Context Protocol (MCP), the open standard that lets AI applications call tools on other systems. In practice it works in two directions: Agentforce agents can call tools on external MCP servers, and outside AI clients can call into Salesforce through MCP servers that Salesforce hosts. Salesforce's Winter '27 developer guide puts it plainly: "Model Context Protocol (MCP) now works in both directions."
That two-way picture is where most explanations go wrong. This guide sorts out what Salesforce documents for each direction, walks through the setup steps for connecting an external MCP server to an agent, and gives admins a governance checklist to run before any MCP connection goes live.
How Agentforce MCP support developed
Salesforce introduced MCP across its platform in June 2025. The developer blog post Introducing MCP Support Across Salesforce said Agentforce would include a native MCP client in pilot with the July 2025 release, and it introduced Salesforce Hosted MCP Servers, MCP support in MuleSoft, and MCP options for Heroku and for developers.
Since then the pieces have moved at different speeds:
- Salesforce Hosted MCP Servers reached general availability. The developer blog post Salesforce Hosted MCP Servers Are Now Generally Available (April 2026) says they are available in "every Enterprise Edition org and above."
- Agents calling MCP tools are described in the Winter '27 developer guide, which says "Your Agentforce agents can call MCP tools, and external AI clients can call your agents as MCP tools," and that you can add AgentExchange MCP server tools to an agent directly in Agentforce Builder, where the Actions modal lists MCP server tools next to standard agent actions.
- The Salesforce DX MCP Server, aimed at developers, is still labeled Beta on Salesforce's Agentforce developer guide MCP page.
Status labels change from release to release, so check the current release notes for your org before you plan around any of these.
Direction one: Agentforce agents calling external MCP servers
This is what most people mean by Agentforce MCP support. An agent in your org connects to an MCP server run by another system, such as an internal service or a partner product, and uses that server's tools as agent actions. Salesforce's developer blog post Extending Agentforce with External MCP Tools via MuleSoft (August 2026) walks through the setup. The MCP for Agentforce requirements in Salesforce Help are Enterprise, Performance, Unlimited or Developer Edition, an agent built in the current Agentforce Builder, and a server that uses Streamable HTTP and, if it needs authentication, OAuth 2.0 client credentials. The same steps apply to any external server:
- Register the server. For a third-party server, go to Setup, enter Agentforce in the Quick Find box, select Agentforce Registry, and click New. Enter a server name, description and URL, choose an authentication method, and click Create and Continue. MuleSoft and Salesforce-hosted servers are managed in API Catalog: in Setup, enter API Catalog in the Quick Find box, then select MCP Servers. Salesforce Help says to create and activate a Salesforce-hosted server there, then register it in Agentforce Registry.
- Allowlist the tools. Review the tools the server offers, select only the ones the agent needs, and click Allow and Continue. Keep the allowlist to the tools the agent actually needs.
- Add the tools to an agent. In Agentforce Builder, choose Add from Asset Library and add the allowlisted tools to the agent as actions. In Winter '27 orgs, the Actions modal in Agentforce Builder also lists AgentExchange MCP server tools.
- Write the agent instructions. Tell the agent when to use each tool and when not to. Instructions guide the agent, but they do not replace access controls, so keep the allowlist narrow.
- Test in a sandbox. Run the conversations your users will have, watch which tools the agent calls, and confirm the external system receives only what it should.
If your company exposes internal APIs through MuleSoft, that same June 2025 announcement says MuleSoft can convert any API or an existing Mule application into an MCP server, which is the route the August 2026 tutorial follows.
Direction two: outside AI clients calling Salesforce through MCP
The second direction runs the other way. An AI client outside Salesforce, such as Claude, connects to an MCP server and works with your org's data and automation. Salesforce's own option is Salesforce Hosted MCP Servers. The Hosted MCP Servers guide says they "give AI agents a secure, governed way to interact with Salesforce data and automation," and that "per-user OAuth 2.0 with PKCE ensures that agents operate with appropriate permissions."
The GA announcement lists prebuilt standard servers for the Agentforce 360 Platform, Tableau Next, Data 360 SQL and more, and says admins can configure custom servers that expose flows, Apex actions and Named Query APIs. You enable a server in Setup, under API Catalog, then MCP Servers. Salesforce's walkthrough Connect Claude with Salesforce Hosted MCP Servers shows the client side: create an external client app with OAuth, require PKCE, and grant the Access Salesforce Hosted MCP Servers (mcp_api) scope.
Winter '27 adds a twist: your Agentforce agents can themselves be exposed as MCP tools to outside clients. The Winter '27 developer guide says only Agentforce agents created with the new Agent Script Builder can be exposed this way, that every call runs as the authenticated user and that Einstein Trust Layer policies still apply.
Where the Salesforce DX MCP Server fits
The third MCP option is for developers rather than admins. The Salesforce DX MCP Server documentation describes it as "an npm package called @salesforce/mcp" that runs locally. You name the orgs it may use with a required orgs setting, and by default it uses only the tools marked GA. It is a good fit for developers working in an IDE with a local project, and less of one for admins who do not work from the command line.
Which kind of Agentforce MCP support do you need?
Start from the job, not the protocol:
- Your Agentforce agents need data or actions from another system. Register that system's MCP server in your org and allowlist its tools. This is the Agentforce MCP client.
- Your team wants to ask Claude or another AI client about Salesforce records, or run existing flows from it. Look at Salesforce Hosted MCP Servers, or at another MCP server that connects to your org.
- Your developers want an AI assistant for code, tests and deployments in their IDE. Look at the Salesforce DX MCP Server.
- Your admins want an AI assistant for configuration work such as fields, flows and permission sets, without building an agent first. That needs an MCP server that can write metadata, covered below.
An admin governance checklist for MCP connections
Every MCP connection gives an AI system a way to act in or around your org. Salesforce's post How to Secure Salesforce Hosted MCP Servers (June 2026) sets out the principles, and they apply to both directions:
- Activate only what you need. Salesforce says "By default, all MCP servers are inactive. Only activate the ones that you intend to expose to agents."
- One app per client. Salesforce strongly recommends creating a dedicated external client app for each MCP client, so you can review and revoke each one on its own.
- Least privilege for the connecting user. "MCP tools run with the same permissions as the user who authenticated." Give that user only the permission sets the job needs. Our daily, weekly and monthly Salesforce admin task checklist includes a quarterly access review that should cover these users too.
- Keep the audit trail useful. Salesforce says actions performed by MCP tools are attributed to the user that connected in audit trails. Avoid shared integration users if you want to know who did what.
- Vet external servers. For the Agentforce client, allowlist individual tools rather than whole servers, and review what each tool can read or change in the other system.
- Test before production. Build and test agent changes in a sandbox, then move them through your normal release process, for example with a Salesforce change set deployed step by step.
MCP for admin work without Agentforce
Agentforce MCP support is about extending agents. Many admins have a simpler need: they want the AI client they already use to help with the configuration backlog. Salesforce's generally available hosted MCP servers can read metadata but can't create or modify it. Cirra AI, the first commercial Salesforce Admin MCP Server, is built for that work. It creates and changes objects, fields, page layouts, validation rules, flows and permission sets from Claude, ChatGPT, Cursor and other MCP clients, and it doesn't need Agentforce or Data 360. One Cirra AI account connects to all your orgs, from production to sandboxes and scratch orgs. Cirra AI runs as the connected Salesforce user, so it can do only what that user can do. Access rules let you set Metadata, User Data, Personal Setup Data and Business Data separately to None, Read Only or Read and Write, and every action is logged in an exportable audit trail. You can compare the options in our guide to choosing and setting up an MCP server for Salesforce.
Frequently asked questions
Does Agentforce support MCP?
Yes. Salesforce says MCP now works in both directions: Agentforce agents can call tools on external MCP servers, and external AI clients can call your agents as MCP tools. Salesforce also offers hosted MCP servers that outside AI clients use to work with Salesforce data and automation.
How do I connect an MCP server to an Agentforce agent?
Register a third-party server in Setup under Agentforce Registry. MuleSoft servers are managed in API Catalog, and Salesforce-hosted servers are activated in API Catalog and then registered in Agentforce Registry. Allowlist the tools the agent needs, then add those tools to the agent from the asset library in Agentforce Builder. Test the agent in a sandbox before production.
What is the difference between the Agentforce MCP client and Salesforce Hosted MCP Servers?
The Agentforce MCP client lets your agents use tools on other systems. Salesforce Hosted MCP Servers work the other way: they let outside AI clients, such as Claude, work with your Salesforce data and automation under per-user OAuth with PKCE.
Which Salesforce editions include Salesforce Hosted MCP Servers?
Salesforce says its hosted MCP servers are available in every Enterprise Edition org and above. Standard servers are inactive by default, so an admin has to enable each one in Setup.
Is the Salesforce DX MCP Server the same as Agentforce MCP support?
No. The Salesforce DX MCP Server is an npm package for developers that runs locally. Salesforce labels it Beta. Agentforce MCP support is about agents in your org using MCP tools and being exposed as MCP tools.
Do I need Agentforce to use MCP with Salesforce?
No. Outside AI clients such as Claude can connect to Salesforce through MCP servers, such as Salesforce Hosted MCP Servers or a third-party MCP server, without you building an Agentforce agent. Check each server's requirements for your org. Agentforce comes in when you want your own agents to call MCP tools.


