How do I connect Claude to my CRM?
Add the CRM's MCP server as a connector with a scoped token. Claude then sees its tools and can read and write within that scope, the same way a person with that token's permissions could.
The short answer
- Connect Claude to a CRM by adding the CRM's MCP server as a connector and issuing a scoped token. Claude then discovers the server's tools and can call them within whatever that token permits.
- If the CRM has no MCP server, the fallback is middleware. A tool like Zapier standing between Claude and the CRM's API, or a custom MCP wrapper a developer writes over the existing API.
- Scope matters more than the connection itself. A token that can read and write everything the CRM contains turns a helpful assistant into a standing risk the moment the wrong instruction is typed.
- Claude's own MCP connector settings live in the client you use, desktop app or CLI, and the exact steps change as Anthropic updates them. Treat the CRM vendor's setup docs as the source for server URL and auth header format.
Why 'just connect it' is not the whole answer
Adding a connector takes a few minutes once you have a server URL and a token. The part that decides whether this was a good idea happens before that: what the token can do. A CRM's MCP server usually mirrors the CRM's own permission model. A token tied to a read-only role behaves like a read-only user. A token tied to an admin role behaves like an admin at the keyboard, except this admin never gets tired and can be told to do something ambiguous at scale.
The second thing that is easy to skip is where the token lives. A token pasted into a desktop client's configuration file sits on that laptop in plain text, unless the client encrypts its config store. If that laptop is lost, shared, or backed up somewhere loose, the token goes with it. None of this is a reason to avoid connecting Claude to a CRM. It is a reason to scope the token narrowly and treat it like any other credential with write access to customer data.
What to check before connecting
Ask each of these before you accept a vendor's 'connect Claude' instructions at face value.
| Question | Why it matters | Where to find the answer |
|---|---|---|
| Does the CRM publish an MCP server? | Determines whether you connect directly or need middleware | The CRM's own integrations or developer docs |
| Where is the token issued? | Determines who can create and revoke it | Usually a Settings or Integrations page inside the CRM |
| What scope options exist? | A token scoped by object, channel and action beats an all-or-nothing token | The token-issuing screen; ask if none is visible |
| Does a write require approval? | Decides whether Claude can act immediately or a person reviews first | The CRM's agent or automation settings, if it has any |
| Can the token be revoked and expire? | Limits the damage window if the token leaks | Same screen as issuance; look for a revoke button and an expiry field |
Setting the scope, not only the connection
Start with the narrowest token that lets Claude do the job you have in mind. If the job is drafting follow-up emails from deal notes, a token scoped to read deals and contacts and write drafts is enough. It does not need permission to delete records or change a deal's stage. Add scope later if a real task needs it, rather than issuing a broad token up front because it is easier.
Disclosure: SalesCrew is our product. Its MCP server exposes 140+ tools across records, inbox, outreach and agent control. Every token is scoped by the same three axes as a human user's access: which channels, which Upwork-style profiles if relevant, and which areas of the app. External sends default to a review queue whatever the token can technically reach. Every call a token makes is written to an audit log with the token named as the actor.
Whatever CRM you use, ask for that same shape. Scoped tokens. A review step on anything that leaves the building. A log you can query afterward. If a vendor's answer to "how do I connect Claude" stops at the connector setup and never mentions scope, that is the gap to push on.
A full-scope token in a desktop client is a credential on a laptop
Questions
- Does this work with Claude desktop and Claude Code the same way?
- The connector setup is the same idea in both. You register the CRM's MCP server URL and a token, and the client discovers its tools. The menu differs between the desktop app and the CLI. Check Anthropic's own MCP documentation for the exact steps in the client you use.
- What if my CRM does not have an MCP server at all?
- Two fallbacks exist. A middleware tool like Zapier can expose CRM actions as steps Claude can trigger indirectly. Or a developer can wrap the CRM's REST API in a small custom MCP server. Both add a layer between Claude and the CRM instead of a direct connection.
- Can Claude see every record in the CRM once it is connected?
- Only what the token scopes allow. A well-built MCP server ties the token to the same permission model as a human user account. A token scoped to one pipeline or one object type cannot read or write outside it, whatever Claude is asked to do.