What an AI agent sees in Salesforce, and how to narrow its permissions
By Asaf Yoshpe, CEO, CloudX
Published: Last tested:
The short answer
An AI agent connected to Salesforce through Salesforce's MCP servers sees what the user who connected it sees, and nothing more. So the first decision is which user it connects through.
When you want the agent to do less than the user, Salesforce recommends a Custom MCP Server: predefined queries (named queries) instead of open access to the data, and defined actions instead of general write permission.
What is MCP in Salesforce
MCP (Model Context Protocol) is an open standard that lets AI tools like Claude or ChatGPT read data and take actions in systems like Salesforce. Salesforce hosts its own MCP servers, and the AI tool connects to them as a specific user.
We already wrote about the difference between Salesforce in Claude and an MCP connection, and how permissions work in NOCA's MCP server. This article picks up from there: exactly what the agent sees in Salesforce, and how to give it less.
Can an AI agent see the CEO's salary?
It's the question that comes up in almost every meeting about connecting AI to a CRM. The answer depends on the user the agent connects through. If an employee has no access to the salary field, the agent connected as that employee doesn't see it either. If the agent connects through a System Administrator user, it sees everything.
Why the agent sees what the user sees
In the MCP servers Salesforce hosts (Hosted MCP Servers), every call runs under the identity of the user who connected. The primary controls are object permissions, field-level security (FLS) and sharing rules. There is no shared service account.
That's good, because everything you've already configured in Salesforce keeps applying. It's also the limitation: if the user's permissions are too broad, so is the agent. A plain-language question quickly exposes permissions nobody noticed.
The standard servers are split by type of action: SObject Reads for reading, SObject Mutations for creating and updating, SObject Deletes for deleting, and SObject All for all of them. There are also servers for Data 360 and Tableau Next. You can already narrow things down by choosing the server: an agent that only needs to answer questions doesn't need the delete server.
How to give the agent less than the user is allowed
Sometimes the user needs broad access and the agent doesn't. A finance manager sees financial data, but it's not obvious their assistant should pull it into every conversation.
According to Salesforce, if you want the agent to do less than the user, you use a Custom MCP Server with named queries. It's an MCP server the admin assembles, with a closed list of tools:
- Predefined queries instead of open access. Instead of the agent writing any query it wants, it gets the queries you defined: which objects, which fields and under which conditions. The agent can't see a field that isn't in any query, even if the user is allowed to see it.
- A defined action instead of general write permission. In Salesforce's example, an investment firm didn't want the agent to act on a client's portfolio. It built a Custom MCP Server with an Invocable Action that only logs a Case, to record the request without giving the agent full write access. That's the principle of least privilege.
This takes planning up front: what the agent really needs, and which actions can be reduced to one defined tool. An organization that exposes the whole API because "it's easier" gives up this layer.
The protective layers around it
A Custom MCP Server is one layer of several. These are the layers Salesforce lists:
- Restrict the user first. Object permissions, FLS and sharing rules. This is the base, and everything else rests on it.
- The agent connects in user context, not through a broad service account. That way the agent's access follows each user's own permissions.
- Restrict who can create External Client Apps. That way an employee can't connect a private AI tool to Salesforce on their own.
- Remove the API Enabled permission from anyone who doesn't need API access.
- Turn on only the MCP servers you need, and turn off the rest in Setup.
- Monitoring. Shield Event Monitoring and Transaction Security, part of Salesforce Shield, to detect and block unauthorized data extraction.
And one more recommendation from our experience: a dedicated user for every recurring use. When the agent runs as a user built for it, with only the permissions it needs, it's easier to check what it did and change its permissions without affecting anyone.
What's changing: Flex Credits and Agentic Identity
Flex Credits on agent calls. Salesforce's official rate card, updated November 3, 2026, lists a new usage type called Headless Platform Interaction, with its rate still marked TBA. According to a Salesforce Ben report on behalf of Salesforce, it covers MCP and API calls by registered agents, in Production only, with 30 days' notice before metering begins, and sandboxes aren't metered. We covered Flex Credits in Salesforce's new editions.
Agentic Identity. According to Salesforce Ben, Salesforce plans to give agents their own credentials and permissions, instead of working under the user's identity. As of November 19, 2026 there is no official Salesforce announcement about it, and the official pages still describe working under the user's identity. If it happens, the question "which user do we connect through" becomes "which permissions do we give the agent". We'll update this article when there is an official publication.
Which official MCP servers does Salesforce have?
| Salesforce Hosted MCP Servers | Salesforce DX MCP Server | |
|---|---|---|
| What it is | MCP servers hosted by Salesforce, for working with data | A local server for developers (the @salesforce/mcp package), with more than 60 tools |
| Who it's for | Users and admins | Developers |
| What the agent sees | What the connected user is allowed to see | A developer tool, not for end-user work |
| Status | Available | Reached end of life on November 2, 2026. Developer tooling is now delivered as Skills |
If you still use the DX MCP Server, it will keep working, but it won't get updates or support. Plan your migration.
On top of these servers there's also Salesforce in Claude, the Anthropic and Salesforce plugin for salespeople. We wrote about it in a separate article.
When you need a paid MCP server
Salesforce's SObject servers work on records: read, create, update and delete. They aren't meant for building the org itself: fields, objects, Flows and permissions. That's where other vendors' servers come in.
Cirra AI
Cirra AI is an MCP server for Salesforce admins that runs as the connected user. According to Cirra, its two main differences are:
- Writing metadata in conversation. Creating objects and fields, building and reviewing Flows, and managing Permission Sets and Page Layouts, straight from the conversation and without a development environment.
- Access control by type of data. Four categories: Metadata, User Data, Personal Setup Data and Business Data, each set to None, Read Only or Read and Write. For example, you can let it build fields without seeing customer data.
According to Cirra, it also has an exportable audit trail and multi-org management from one place. Running existing Flows is marked "coming soon", and according to Cirra's comparison, Salesforce's Hosted MCP already supports it. Cirra has Free, Professional and Team plans (pricing).
Other servers
| Server | What it does, according to the vendor |
|---|---|
| CData | An MCP server that understands the Salesforce data model, delivered as a remote endpoint through CData Connect AI |
| Coupler.io | A hosted MCP server that syncs CRM data on a schedule so Claude or ChatGPT can answer pipeline questions. Read-only |
| SyncHub | An MCP server connecting Salesforce and more than 70 other cloud services. Read-only |
| MuleSoft | An MCP Connector that lets Mule apps act as both MCP server and client, exposing APIs, connectors and apps to AI |
MCP or Agentforce?
Agentforce is Salesforce's AI inside Salesforce. An MCP server brings the AI tools your team already uses, like Claude, to Salesforce data. You can use both. Which agent fits which task? We have a webinar on that.
How we do it with clients
CloudX has an AI procedure, and this is what it means in practice:
- A sandbox with synthetic data by default. Work with AI tools starts in a test environment without real data.
- A signed approval addendum before connecting AI to a client's environments. The client signs before we connect AI tools to their environments, including MCP connections, and chooses which environments and for how long.
- Signing off on a connection isn't a license for an agent to run on its own in production. Connecting for a piece of work is one thing. An autonomous agent in production is a separate decision.
- Cirra AI is approved in our AI procedure, and our team uses it for metadata and permission checks.
Before connecting AI to Salesforce: a checklist
- Decide which user the agent connects through, and check what that user actually sees.
- Clean up overly broad permissions before connecting, not after.
- Turn on only the MCP servers you need.
- For sensitive data, build a Custom MCP Server with defined queries and actions.
- Restrict who can create External Client Apps, and remove API Enabled from anyone who doesn't need it.
- Start in a sandbox, with synthetic data.
- Set up monitoring, and check from time to time what the agent did.
Before you connect, we'll check what the user your agent connects through actually sees, and close what isn't needed. Talk to us.
FAQ
What is MCP in Salesforce?
MCP (Model Context Protocol) is an open standard that lets AI tools like Claude or ChatGPT read data and take actions in Salesforce. Salesforce hosts its own MCP servers, and the AI tool connects to them as a specific user.
Does Salesforce have an official MCP server?
Yes. Salesforce hosts MCP servers (Hosted MCP Servers) for working with data: reading, creating, updating and deleting records, plus servers for Data 360 and Tableau Next.
What does an AI agent see in Salesforce? Can it see fields I don't have access to?
Through Salesforce's MCP servers, the agent sees what the user who connected it sees, and nothing more. Every call runs under the user's identity, with their object permissions, field permissions and sharing rules.
How do you give an AI agent fewer permissions than the user?
With a Custom MCP Server: a server the admin assembles with a closed list of tools. Instead of open access you define queries up front (named queries), and instead of general write permission you define actions, like an Invocable Action.
How do you stop employees from connecting a personal AI tool to Salesforce?
Salesforce recommends restricting External Client Apps creation to authorized personnel only, and removing the API Enabled permission from anyone who doesn't need API access.
How do you connect Claude to Salesforce?
A Salesforce admin turns on the MCP servers you need, and each user connects from Claude with their own Salesforce user. For salespeople there's also the Salesforce in Claude plugin, with an Anthropic setup guide.
Does Salesforce charge for MCP?
Salesforce added a usage type called Headless Platform Interaction to the Flex Credits rate card, and the rate is not yet published. According to a Salesforce Ben report, it will apply to calls by registered agents in Production only, with 30 days' notice.
What is the difference between Salesforce Hosted MCP and Cirra AI?
Salesforce's SObject servers work on records. According to Cirra, its server adds writing metadata in conversation (fields, objects, Flows and permissions) and access control by type of data. Running existing Flows is marked "coming soon" at Cirra.
What happened to the Salesforce DX MCP Server?
It reached end of life on November 2, 2026. Existing installs keep working, but without updates or support. Developer tooling is now delivered as Skills.
Is it safe to connect AI to Production?
It depends on permissions, the servers enabled and monitoring. We start in a sandbox with synthetic data, and connect to a client's environments only after a signed approval addendum. Even then, signing off on a connection isn't a license for an agent to run on its own in production.
Sources and date checked
These details were checked on November 19, 2026. The vendor's site is always the authoritative source, so check there before deciding.
- Salesforce: the trust model of agentic architecture (Headless 360)
- Salesforce: Design an AI-native architecture with Headless 360
- Slack: Salesforce MCP servers
- Salesforce: Flex Credits rate card, November 3, 2026
- Salesforce Ben: Flex Credits for agentic MCP and API calls
- Anthropic: Salesforce in Claude
- GitHub: Salesforce DX MCP Server
- GitHub: DX MCP Server end-of-life notice
- Cirra AI: Salesforce MCP Server
- CData: MCP Server for Salesforce
- Coupler.io: Salesforce MCP
- SyncHub: Salesforce
- MuleSoft: MCP Connector