Connect Your CRM to Claude or ChatGPT: An MCP Setup Guide

To connect your CRM to Claude or ChatGPT, you point the assistant at the vendor's Model Context Protocol (MCP) server, approve an OAuth login, and from then on the AI can read and write your pipeline in plain language. No scripts, no glue code. Six of the major CRMs already ship an official MCP server, so for most teams this is a five-minute setup rather than an engineering project.

This is the practical companion to the 2026 MCP readiness report, which scored which tools are agent-ready. This guide covers the next question: once a tool is ready, how do you actually plug it in, and what can you do once it is connected.

What MCP is, in one paragraph

MCP is an open standard that lets an AI assistant talk to an outside tool through a defined set of actions. A CRM's MCP server exposes things like "search contacts", "update a deal", or "log a call" as tools the model can call. You connect once, grant access, and the assistant treats your CRM like a native capability. The difference from a chatbot that "knows about" your CRM is that an MCP connection acts on live data: it reads the actual record and writes the actual change.

If you want the deeper background on why this matters for a sales stack, how to build an AI-native sales stack covers the strategy. This post stays hands-on.

What you need before you start

An AI assistant that supports remote MCP connectors. As of June 2026 that includes Claude (web, desktop, and Claude Code), ChatGPT (via connectors and developer mode), Cursor, and a growing list of others.
An admin or appropriately permissioned login to your CRM. MCP respects your existing permissions, so the assistant can only touch what your account can touch.
Two minutes to approve an OAuth screen. Most official servers are hosted by the vendor, so there is nothing to install.

A note on permissions before you connect anything: an MCP server can write to your CRM. Start with read-only intent, watch what the assistant proposes, and approve writes deliberately until you trust the workflow. Treat it like giving a new rep edit access.

The general flow

Every official CRM connection follows the same three steps:

Find the server URL. The vendor publishes it. We list the exact endpoint for each CRM below.
Add it as a connector. In Claude, that is the connectors or MCP settings screen, or claude mcp add in Claude Code. In ChatGPT, it is the connectors panel. The assistant asks for the server URL.
Approve OAuth. A vendor login screen opens, you confirm access, and the tools appear. Ask the assistant "what can you do in my CRM now" to confirm the connection is live.

The rest of this guide is the per-tool detail: the server, what it can do, and where it shines.

CRM by CRM

Attio

Attio is the cleanest connection in the catalog. It ships a first-party hosted MCP server at mcp.attio.com over OAuth, on top of a REST API that reaches every record and attribute. Because Attio was built API-first, the assistant can read and write the entire pipeline with no gaps. Works with Claude, ChatGPT, Gemini, and Zapier or Make. At an AI-stack-fit score of 88, it is tied for the top of the leaderboard. If you are choosing a CRM with agents in mind, this is the reference implementation.

HubSpot Sales Hub

HubSpot ships an official remote MCP server that lets an assistant read, search, and act on contacts, deals, and engagements over OAuth. The CRM REST API underneath is mature and uses date-based versioning, and HubSpot's own Breeze agents run on the same plumbing, so you are connecting to well-trodden infrastructure. Works with Claude, ChatGPT, Gemini, and Zapier or Make.

Close

Close ships a vendor-supported MCP server at mcp.close.com (HTTP Streamable, OAuth 2.0), plus a complete REST API and webhooks. The combination means an assistant can log activity, update leads, and trigger sequences natively, which is exactly what you want from a CRM built for high-velocity outbound. It also scores 88. Works with Claude, ChatGPT, and Zapier or Make.

Salesforce Sales Cloud

Salesforce ships official hosted MCP servers with SObject operations that read and write records under field-level security, connectable from Claude, ChatGPT, and Cursor. The platform also exposes REST, SOAP, and GraphQL APIs. The catch is governance: setup is heavier and you will likely route this through your admin. If you are already on Salesforce, the security model is a feature, not a bug, because the assistant inherits your field-level permissions.

Zoho CRM

Zoho CRM ships pre-built official MCP servers covering data query, record operations, customization, and workflow, and they install into Claude Desktop, Cursor, and VS Code. Everything the servers do not cover is reachable through the REST API v8. Connects to ChatGPT through the API and into Zoho's own Flow automation.

monday CRM

monday CRM ships an official MCP server (the mondaycom/mcp project, preinstalled on all accounts at no extra cost) that supports CRM actions like creating leads and deals. The platform API is GraphQL-only at a single endpoint, so any deeper automation builds against that. Works with Claude, ChatGPT, and Zapier or Make.

Two more CRMs worth a look if you want a lighter setup: Pipedrive and Folk both sit on the CRM roundup and rate well for API access. Compare all of them on the best CRM page.

Beyond the CRM: the rest of the stack worth connecting

Your CRM is the core, but an agent that can only touch the CRM is half-blind. The highest-leverage additions:

Prospecting and enrichment

Clay ships a 21-tool GTM MCP server at clay.com/mcp for Claude, ChatGPT, and Codex, on top of its enterprise REST API. Connect it and the assistant can build and enrich lists, then hand them to your CRM in the same conversation.
Apollo.io ships its own MCP server with around 45 tools and OAuth setup in Claude, ChatGPT, and Perplexity, available on all plans. Good for search and enrichment, though rate limits bite on large pulls.

Both lead the prospecting roundup.

Meetings and call intelligence

Fireflies.ai ships an MCP server at api.fireflies.ai/mcp so an assistant can search calls and pull transcripts on demand, backed by a clean GraphQL API.
Gong ships an official MCP server plus a documented REST API exposing calls, transcripts, and deal data.

Getting the transcript into the assistant is the whole point of this category, which is why it tops the readiness report. See the full list on the conversation intelligence roundup.

What to actually do once it is connected

A connection is only worth setting up if it saves you real time. The patterns that pay off fastest:

Morning pipeline review. "Summarize every deal that moved stage yesterday and flag the ones with no next step." The assistant reads the CRM and writes you a briefing.
Post-call updates. Pull the transcript from Fireflies or Gong, draft the CRM note, and update the deal fields, in one prompt.
List building to outreach. Enrich a target list in Clay, push it to the CRM, and draft the first-touch sequence.
Hygiene sweeps. "Find contacts with no email and a deal over 30 days old." The boring maintenance an agent is actually good at.

For seven worked examples of running a GTM motion this way, see how to run your GTM stack with Claude Code.

Common gotchas

The tool has an API but no MCP. You can still connect it through a bridge like Zapier or Make, or a community MCP server, but the experience is rougher and you own the maintenance. The readiness report flags which tools are in this bucket.
A vendor markets itself as "AI" but exposes nothing. This is most common among AI SDR products, the lowest-scoring category in the catalog. A product built to be an agent often has no interest in being driven by yours. Check the MCP and API line, not the homepage.
Permissions surprise you. If the assistant cannot see a record, it is usually your CRM permissions, not the connection. MCP inherits them by design.
Writes feel risky. Keep a human in the loop on bulk changes until you trust the pattern. Start with reads.

How to choose, if you are still deciding

If agents will run more of your pipeline every quarter, weight the connective tissue first. Your CRM and prospecting layer get touched by every workflow, so a high AI-stack-fit score there compounds. Attio and Close lead the CRMs at 88, with Salesforce, Zoho, HubSpot, and monday all shipping official servers close behind.

The full, sortable scores live on the AI-stack-fit leaderboard, and the data behind every number is in the 2026 MCP readiness report. Server URLs and connection details reflect what Chief Revenue Buddy verified in June 2026 and are refreshed as vendors ship updates.