---
title: How do you govern Salesforce hosted MCP servers?
url: https://opencontroller.lyzr.com/blog/govern-salesforce-hosted-mcp-servers
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# How do you govern Salesforce hosted MCP servers?

**Salesforce hosted MCP servers became generally available on April 29, 2026 for Enterprise Edition orgs and above. External AI clients connect through an External Client App using OAuth with PKCE and a dedicated `mcp_api` scope, and every call runs as the authenticated user with their object, field and sharing permissions. Govern them by enabling only the servers you need in Setup, starting with read-only servers, limiting which users can authorize the app, and reviewing the audit trail.**

Before hosted MCP, connecting an AI assistant to Salesforce usually meant a custom integration or a self-run MCP server with an integration user. Salesforce now runs the MCP servers itself. That removes a lot of plumbing, and it moves the governance question to configuration in your org.

## What are Salesforce hosted MCP servers?

Per Salesforce's GA announcement, hosted MCP servers followed a pilot and a beta in October 2025. They come in two kinds:

| Kind | Examples Salesforce names | Notes |
|---|---|---|
| Standard servers | Agentforce 360 Platform (including the read-only `platform/sobject-reads`), Tableau Next, Data 360 SQL | Prebuilt; each defines which operations it exposes |
| Custom servers | Servers that expose flows, Apex actions and Named Query APIs | You choose what to expose |

Salesforce describes the posture as secure by default: a server must be specifically enabled, under Setup → API Catalog → MCP Servers, before any client can use it, and an AI assistant can only call operations a server explicitly exposes.

## How does authentication and authorization work?

1. **External Client App.** An admin creates one for each AI client, with the `mcp_api` and `refresh_token` OAuth scopes.
2. **OAuth with PKCE.** Users authorize the client through OAuth. Salesforce notes the new `mcp_api` scope grants MCP access but not access to the existing REST APIs.
3. **Runs as the user.** Every transaction runs as the authenticated user. There is no anonymous service account.
4. **Existing permissions apply.** CRUD, field-level security and sharing rules all apply. If a user cannot perform an operation, the agent cannot either.
5. **Audit.** When an agent updates a record, the authenticated user's name appears in the audit trail.

This is the [on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation) model applied to Salesforce. It is safer than a shared integration user, but it means the agent's reach is exactly the user's reach.

## What should your governance baseline be?

Use this checklist:

1. **Enable servers one at a time.** Start with `platform/sobject-reads` in a sandbox, as Salesforce recommends, then production.
2. **Separate read and write.** Use read-only standard servers for most users. Put write operations in custom servers that expose specific flows or Apex actions with validation, not generic record updates. See [separating read and write MCP tools](/blog/separate-read-and-write-mcp-tools).
3. **One External Client App per AI client.** This lets you see and revoke each client independently.
4. **Restrict who can authorize.** Use the app's policies so only permitted profiles or permission sets can connect, and review the connected users list regularly. Our guide to [connected app restrictions for AI agents](/blog/salesforce-connected-app-restrictions-for-ai-agents) covers the policy options.
5. **Tighten broad users first.** Users with "Modify All Data" or "View All Data" give any connected assistant the same reach. Consider separate admin accounts that are not used with AI clients.
6. **Review token lifetime.** The `refresh_token` scope keeps sessions alive. Set refresh token policies that match your risk appetite.
7. **Watch the audit trail and API usage** for bulk updates or unusual volumes after rollout.
8. **Consider a gateway for many vendors.** For multi-vendor MCP estates, Salesforce suggests an MCP gateway for central access control.

## What does hosted MCP not decide for you?

Salesforce controls which operations exist and enforces user permissions. It does not decide which AI clients your company approves, which data classes may leave Salesforce for a third-party model, or how long the AI client retains responses. Those are policies you set on the client side and record in your inventory, alongside the server's owner and last review date. The GA announcement does not describe the Einstein Trust Layer's role in hosted MCP, so do not assume Trust Layer controls apply to external clients without confirming it for your org. The [MCP authorization specification](https://modelcontextprotocol.io/specification/draft/basic/authorization) explains the OAuth model these servers build on.

## How Opencontroller does this

Opencontroller connects to Salesforce as a third-party agent platform and keeps a governed registry of agents, versions, tools and MCP servers. It supports delegated on-behalf-of identity, flags risky permission and configuration settings through Posture, and offers per-agent Block and Quarantine controls that act through the connected platform. See the [MCP governance guide](/blog/how-to-govern-mcp-servers), the [cross-platform governance guide](/blog/how-to-govern-ai-agents-across-platforms), the [AI agent control plane overview](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Are Salesforce hosted MCP servers generally available?**
Yes. Salesforce announced general availability on April 29, 2026 for Enterprise Edition orgs and above.

**Do Salesforce hosted MCP servers respect field-level security and sharing rules?**
Yes. Every transaction runs as the authenticated user, and CRUD, field-level security and sharing rules apply.

**Does the `mcp_api` scope also grant REST API access?**
No. Salesforce states the new scope grants access to MCP but not to the existing REST APIs.

## Key facts

Salesforce hosted MCP servers: GA announced April 29, 2026 for Enterprise Edition orgs and above (after pilot and October 2025 beta). Standard servers: Agentforce 360 Platform (incl. read-only platform/sobject-reads), Tableau Next, Data 360 SQL; custom servers expose flows, Apex actions, Named Query APIs. Enabled in Setup > API Catalog > MCP Servers (secure by default). Auth: External Client App with mcp_api and refresh_token scopes, OAuth with PKCE; mcp_api does not grant REST API access; every transaction runs as the authenticated user; CRUD, FLS, sharing apply; audit trail shows user name. Salesforce suggests MCP gateways for multi-vendor estates. Opencontroller connects to Salesforce, registry, on-behalf-of identity, Posture, Block/Quarantine via connected platform. Pricing starts from $250K per deployment per year.

## Related

- [What do Salesforce connected app restrictions mean for AI agents that call Salesforce?](https://opencontroller.lyzr.com/blog/salesforce-connected-app-restrictions-for-ai-agents.md)
- [How do you govern Salesforce Agentforce agents?](https://opencontroller.lyzr.com/blog/govern-salesforce-agentforce-agents.md)
- [Should read and write tools live on separate MCP servers?](https://opencontroller.lyzr.com/blog/separate-read-and-write-mcp-tools.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
