Salesforce MCP Permissions: What Your AI Can Actually See
Can Claude See Salesforce Data You Can't? How MCP Permissions Work
The first question I hear when a team wants to connect Claude to Salesforce isn't about features. It's from whoever owns security: "So the AI can see everything now?"
No. But the answer depends on who connects it, and that's where teams get caught.
Short answer: A Salesforce MCP server never sees more than the person who connected it. Each user signs in with their own Salesforce login, and every request runs with that user's object permissions, field-level security, and sharing rules. The risk isn't the AI going around your permissions. It's an admin connecting with an admin's access.
Does a Salesforce MCP server bypass user permissions?
No. Salesforce stays the authority on its own data.
When someone connects Claude or ChatGPT to a Salesforce hosted MCP server, they sign in through Salesforce. Their individual user account is tied to that session. From there, Salesforce's security guide says every tool runs with that user's permissions, including:
- Object-level access (create, read, update, delete)
- Field-level security
- Record sharing rules
So if a rep can't see the Finance team's margin field in Salesforce, they can't see it through Claude either. Same records, same fields, same rules you already maintain.
Salesforce also doesn't let you connect hosted MCP through a shared service account. Every person authenticates as themselves. Salesforce calls the shared-account approach an anti-pattern, and I agree.
That's about to change. Salesforce has announced agent registration, targeted for November 2026, with no launch date published as of mid-November. Admins will give each AI agent its own identity with a narrower permission set, instead of letting it borrow the signed-in user's access. Once Salesforce gives notice that registration is available, any use of a Salesforce MCP server will need it within three months. The rules in this post apply today. Expect the model to shift from "runs as you" to "runs as an agent you've scoped."
What happens when an admin connects?
The AI gets an admin's view. That's the part people miss.
Mitch Carlson, founder of UpNorth Software, a Salesforce consulting practice, described it simply when we talked: if a system administrator who can see everything in the org connects, then everything they can access in Salesforce, they can access through their AI. If a rep with limited access connects, the AI stays inside that rep's view.
The permission model works exactly as designed. That's why it matters who connects, and to which server.
An admin on a full-access server can ask Claude to update or delete records across the whole org. One badly worded prompt can change data your forecast depends on. Admins should start on the read-only server like everyone else.
Read-only, write, or delete: does the server matter?
Yes. The server is your second layer of control.
Permissions decide which records a user can touch. The server you activate decides which actions are on the menu at all. Salesforce offers standard servers for read-only, create and update, delete, and full access. If you only activate the read-only server, nobody's AI can change data through it, admin or not.
Think of it as two locks. The user's permissions are one. The server's menu is the other. The server lock is org-wide, though: once a write server is on, it's on for everyone who can connect.
Who can connect to your Salesforce MCP server?
By default, anyone in your org. That surprises people.
Once you create the external client app that connects AI tools to your MCP servers, any user can authenticate through it unless you restrict it. Salesforce lets you limit access to users with specific profiles or permission sets. For most teams, a dedicated permission set for MCP users is the cleanest way to control rollout.
How do you see what the AI did?
Through Event Monitoring. Every action an MCP tool takes is attributed to the user who connected, so your audit trail stays intact.
In Setup, open the Event Log File Browser and filter the event type to API Total Usage. MCP traffic shows up with the API client category SALESFORCE_HOSTED_MCP, along with the user and the objects touched. That gives you a way to spot errors, odd access patterns, or someone using write tools you didn't expect.
Six settings to check before you roll it out
1. Activate only the servers you need.
All servers start inactive. Turn on read-only first. Add write access later, once you've seen how people use it. Once you activate a write server, every connected user can write wherever their Salesforce permissions allow, so control who writes through their permissions and who can connect, not the server.
2. Use one external client app per AI tool.
Salesforce recommends separate apps for Claude, ChatGPT, and Cursor. It makes access easier to control and activity easier to audit.
3. Use the MCP scope, not full API access.
Give the app the "Access Salesforce hosted MCP servers (mcp_api)" scope. It opens MCP without opening every Salesforce API.
4. Restrict access to a permission set.
Otherwise every user in the org can connect. Start with a pilot group.
5. Shorten the refresh token lifetime.
The default token lasts a year. Shorter is safer in production. If you ever need to cut access fast, you can revoke tokens from OAuth Usage in Setup.
6. Review the logs on a schedule.
Without the Event Monitoring add-on, Salesforce keeps the API Total Usage log for only 24 hours, so export it regularly or license Event Monitoring for longer retention. A regular look at MCP activity in Event Monitoring catches problems before they show up in a board report.
Frequently asked questions
Can Salesforce MCP access data a user can't see?
No. Every request runs with the connected user's object permissions, field-level security, and sharing rules.
Can I connect Salesforce MCP with an integration user?
Not today. Each person signs in with their own account, and Salesforce doesn't offer a shared service account option for its hosted servers. Agent registration, which Salesforce has announced but not yet launched, is the planned route for giving an AI agent its own identity.
How do I make Salesforce MCP read-only?
Activate only the read-only standard server (sobject-reads) in Setup under API Catalog, then MCP Servers.
Are Salesforce MCP actions logged?
Yes. Event Monitoring records MCP activity and attributes it to the user who connected.
The honest takeaway
Salesforce got the permission model right. The AI inherits your existing security instead of inventing its own. That also means any gaps in your profiles and sharing rules now show up in a chat window, faster than they used to. If your permissions have drifted over the years, fix that before you connect anything. Start read-only, start small, and watch the logs.
New to the basics? Start with What Is a Salesforce MCP Server?
If you'd like a second set of eyes on your permissions before rollout, that's work we do at Sales Tempo.
Salesforce MCP permissions · Salesforce security · sharing rules · field-level security · Model Context Protocol · RevOps
Sales Tempo · salestempo.io
