What Is a CRM MCP Connector and How Do You Run Your CRM From Claude in 2027?
A CRM MCP connector lets Claude read and update your CRM within your permissions. How it works, how to add one and four questions to ask first.
Contents7 sections
In a test published in March 2025, the Tow Center for Digital Journalism found that eight AI search tools, asked to name the source of news excerpts, gave wrong answers to more than 60 percent of 1,600 queries. Most answered "with alarming confidence". Run your CRM from Claude through an MCP connector, and a confident mistake can land in a deal. The connector lets Claude read and change your CRM as you, and you decide which changes go through.
This is for the owner who still sells, with Claude open all day and the CRM opened only when something forces it. You'll know what a connector is, how to add one and what to ask before you switch it on.
How we put this together. This rests on the Model Context Protocol's documentation and Anthropic's Claude Help Center, both read on 19 November 2026; the four checks are ours.
What is a CRM MCP connector?
A CRM MCP connector links an AI assistant such as Claude to your CRM, so it can read and change records signed in as you. We'd let it read freely and approve every change for the first month, because Anthropic's Claude Help Center says to click "Allow always" only for a tool you trust to run unsupervised.
MCP stands for Model Context Protocol. In its documentation, it is "an open-source standard for connecting AI applications to external systems", likened to "a USB-C port for AI applications". The same page names Anthropic's Claude and OpenAI's ChatGPT among the applications that support it.
A CRM connector is that port for one system. The CRM's side runs an MCP server, which offers the CRM's actions as tools: find a contact, read a deal, add a note, move a deal to its next stage. The assistant calls those tools when your question needs them.
Two things it isn't. Your records don't move into Claude. They stay in the CRM, and Claude asks the server for the few it needs each time you ask something. And it isn't the assistant's memory: a deal Claude mentions in a chat changes nothing in the CRM until a write tool saves it.
How does a CRM MCP connector work? Read first, write on approval
You add the CRM's server address in Claude, sign in to the CRM, and ask in plain words. Claude picks one of the server's tools and asks you before using it; you allow it, and the CRM answers or changes the record. Read first, write on approval means you allow reads easily and look hard at writes.
- The CRM offers an MCP server. It lists what an assistant may do, tool by tool. Whoever built it decides which tools exist and whether any of them write.
- You add it to Claude as a connector, with the server's URL (the clicks are in the start section below).
- You sign in to the CRM. Connecting typically uses OAuth, so Claude acts for you "without Claude ever seeing your actual password". It reaches what your account reaches.
- You ask. For example "which deals haven't moved this month?" or "add a note to the Berger account that they want a quote in March" (both invented).
- Claude asks before it uses a tool. You allow the call once, or choose "Allow always" for that tool.
- The result comes back. The server does the work inside the CRM, and Claude tells you in the chat what it found or saved.
A wrong answer costs you a second look at the CRM screen. A wrong write changes a deal stage or a contact that a colleague will trust tomorrow without checking. So once a few answers have matched what the CRM shows, let the read tools run, and keep every write tool on ask-each-time.
See Saga, the CRM that comes with HAL, gethal.ai's AI sales agent, run from Claude. See Saga run from Claude
What changes compared with opening the CRM?
You stop clicking through screens and start asking in sentences. Claude does the clicking through the server's tools, signed in as you, and stops to ask before each one unless you've allowed it always. Your CRM keeps its records and its history. Two things move: who presses the buttons, and how you stop them.
| Opening the CRM | Through a connector in Claude | |
|---|---|---|
| How you ask | Menus, filters and search boxes | A sentence in the chat |
| Who clicks | You | Claude, calling the server's tools |
| What it can reach | Whatever your login shows | The server's tools, within your account's access |
| Approving a change | Pressing Save is the approval | Claude asks before each tool call, unless you chose "Allow always" |
| Setting it up on a Team or Enterprise plan | Set in the CRM itself | An Owner adds the connector in Claude first |
| How you stop it | Log out | Disconnect the connector, which revokes its permissions |
The last three rows are the ones the CRM screen never asked you to think about. On the screen, Save happens because you pressed it. In Claude, a call can run on its own once you've chosen "Allow always", and the server decides which actions exist at all.
The top rows are why people want one: you ask "who did we last speak to at the Linz account?" (invented) in the chat where you were already working, and the CRM's screens stay for reading a whole account before a meeting.
See Saga's accounts and deal board in a demo. See Saga in a demo
Is it safe to connect a CRM to Claude?
It can be, if you control three things: whose account the connector uses, which tools may run without asking, and which server you added. Claude's custom connectors usually sign in over OAuth, ask before using a tool and can be disconnected at any time. The risk sits in servers you don't know and writes nobody looked at.
Permissions: it acts as you
The sign-in in step 3 means Claude reaches what your CRM account reaches. If you're the CRM's admin, that can be every account and every deal. Where your CRM allows it, connect with a user that sees only what you work on.
Approval: what "Allow always" hands over
Once you click "Allow always", that tool runs without stopping to ask. Anthropic's advice, quoted in the first answer above, is the line to hold: keep it for tools that only read.
Revoking: one click in Claude
In Anthropic's words, you can "revoke these permissions at any time by disconnecting the connector". Do it the day you stop using a connector, or the moment a tool does something you didn't ask for.
Unknown servers: hidden instructions
Anthropic warns that "Malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions". A server sits between Claude and your customer records, so add one only from your CRM's maker or a provider you can name and check.
Confident mistakes
In the Tow Center's test (6 March 2025), even the tool with the fewest errors answered 37% of queries incorrectly. It measured AI search tools finding news sources, not CRM work. In your first week, check a few of Claude's answers against the CRM screen.
What does a connector change for the owner and a small team?
It moves quick questions and short updates into the assistant you already have open. The owner asks about the pipeline between calls; the salesperson adds a note while the call is still fresh. Logging emails and meetings is a different job: a CRM that updates itself does that, and the connector covers what only you know.
BCG wrote in 2019 that salespeople who see strict data rules as extra admin may skip logging client interactions, which it calls "passive disobedience". BCG was writing about large sales teams. In a team of three, the same habit shows up as the update that waits until Friday because it means opening another tool. A connector puts that update where you already are, for example (all invented):
- Questions about the pipeline: "which deals have no next step?"
- One-line updates: "book a call with Anna at Berger next Tuesday".
- A check before a call: "what's open on the Berger deal?"
How to word a one-line update so the CRM gets every part right is in our guide to updating your CRM by chat. Where the hours go when a team keeps the CRM current by hand is in how much time updating the CRM takes.
A self-updating CRM puts emails, replies and meetings in the account's timeline without anyone typing them. That leaves the connector for what only you know: what the buyer said on the phone, and when you promised to call back. Large changes, such as moving 200 contacts to a new owner, stay on the CRM's own screens, where you see the whole list before you press the button.
How do you start, and what should you ask any provider?
Get your CRM's MCP server address, add it in Claude under Customize > Connectors, sign in, and ask one read-only question first. Before you do, put four questions to the provider: what it can read, what it can write, whose permissions it uses and how you stop it. Written answers to all four tell you what you're switching on.
| Step | Where | What you get | Cost |
|---|---|---|---|
| 1. Get the server's URL | Your CRM provider | The address of its MCP server | Set by your CRM provider |
| 2. Add the connector | Claude, Customize > Connectors | The CRM in your list of connectors | Free plan: one custom connector; also on Pro, Max, Ultra, Team and Enterprise |
| 3. On Team or Enterprise, an Owner adds it first | Claude, by your organisation's Owner | The connector available to add | Within your Claude plan |
| 4. Sign in | Your CRM's sign-in, usually OAuth | Claude acting for you without your password | Your Claude and CRM plans |
| 5. Ask one read-only question | The chat | A first answer to check against the CRM screen | Your Claude and CRM plans |
| 6. Allow writes one by one | Claude's prompt before each tool call | Every change seen before it's written | Your Claude and CRM plans |
Four questions to ask any provider
| Check | Ask the provider | A good answer |
|---|---|---|
| Read | Which records can the server read? | A written list of its tools, by record type |
| Write | Which tools write, and can I switch them off? | The write tools named one by one, and a read-only option |
| Permissions | Does it sign in as me, with my own CRM access? | Yes, over OAuth, with your user's rights, not a shared admin key |
| Stop | How do I cut it off, and who runs the server? | Disconnecting in Claude ends it, and the server is run by a company you have a contract with |
Use HAL when you want your CRM in the assistant you already use: Saga, the CRM that comes with HAL, is an MCP connector, so you run it from Claude or any other MCP client.
See Saga work as an MCP connector in a demo. See Saga's connector
Who this is for
- The owner who still sells: the first answer and the start section, so you can ask about the pipeline from Claude this week and keep every change on approval.
- The salesperson next to the owner: the small-team examples, for notes and calls added while the conversation is still fresh, without opening the CRM between calls.
- Whoever owns your Claude Team account: the safety section and the four questions, since an Owner adds the connector before anyone else on the team can use it.
- Whoever keeps the CRM tidy: the approval rule and the comparison table, so every change made from Claude is one somebody looked at first.
FAQ
What is an MCP connector?
An MCP connector links an AI assistant such as Claude to another system, such as a CRM, through that system's MCP server. The server offers actions as tools, and the assistant calls them when your request needs them.
What is MCP in simple terms?
MCP, the Model Context Protocol, is an open standard that lets AI applications connect to outside data, tools and workflows. Its own documentation compares it to "a USB-C port for AI applications". Claude and ChatGPT both support it.
Can Claude connect to my CRM?
Yes, if your CRM offers an MCP server, or a provider you trust runs one for it. You add the server's URL in Claude as a custom connector and sign in to the CRM.
Is it safe to connect a CRM to Claude?
It is when you control the account, the approvals and the server. Claude signs in for you, usually over OAuth, without seeing your password, asks before it uses a tool, and loses access when you disconnect it. Keep write tools on approval, and add servers only from a provider you can name, since Anthropic warns that malicious servers can hide instructions.
How do you add a custom connector in Claude?
Open Customize > Connectors in Claude, add a custom connector and paste the MCP server's URL. Then sign in to the service, usually through OAuth. On Team and Enterprise plans an Owner adds the connector first.
What is a CRM MCP server?
It's the program on the CRM's side that offers the CRM's actions, such as reading a deal or adding a note, as tools an AI assistant can call. Whoever builds it decides which tools exist and whether any of them write. Ask the provider for that list before you connect.
Run your CRM from the assistant you already use
Start with one read-only question this week and keep every write on approval. To watch Saga run from Claude, book a demo. Saga is free with every Outreach plan, so HAL's prices cover it.
Keep reading: What is a self-updating CRM · How to update your CRM by chat · Time spent updating the CRM


