How to Connect Claude to Your CRM (Step by Step)

Connecting an AI assistant like Claude to your CRM used to require custom code. With the Model Context Protocol (MCP), it is now a one-time authorization. This guide walks through the whole process — what you need, how the connection works, the first prompts to try, and how to keep agent access safe.

What you will need

Three things:

• An MCP-compatible client. Claude Desktop is the most common starting point, but any client that speaks MCP works, including Claude Code and the claude.ai connectors.
• A CRM with an MCP server. Your CRM needs to expose an MCP endpoint. Kantos includes one on every paid plan.
• An account with permission to authorize access. Connecting an agent grants it access to CRM data, so you will sign in and approve the connection.

Step 1: Choose your MCP client

Decide where you want the agent to live. If you want a desktop assistant that can also touch local files, Claude Desktop is a good default. If you work in a terminal or editor, Claude Code is a strong fit. The connection process is the same idea in each: tell the client where the CRM's MCP server is, then authorize.

Step 2: Point your client at the CRM's MCP server

Every MCP client has a place to register servers — usually a settings file or a connectors screen. You add an entry that names the server and gives its URL. For Kantos, that URL is the hosted MCP endpoint shown in your account and in the MCP server and REST API reference. You do not need to paste API keys into the config for the agent flow; authorization happens in the next step.

Step 3: Authorize with OAuth

The first time the agent tries to use the server, it starts an OAuth authorization. A browser window opens, you sign in to your CRM, and you approve the access. This is the same pattern you have used to connect any app to Google or GitHub. Behind the scenes the agent receives a scoped token — it can do what you approved and nothing more, and you can revoke it at any time. After you approve, the client lists the CRM's tools as available, and the agent can act on your data.

Step 4: Try your first prompts

Start with read-only requests to build confidence, then move to actions. Good first prompts:

• "List the leads created this week and summarize where each one stands."
• "Show me the open deals worth more than 10,000 and who owns them."
• "Find the contact named Maria Gomez and tell me the history of our last three interactions."

Once you trust the reads, try writes:

• "Move the Acme deal to the proposal stage and add a note that I sent pricing today."
• "Draft a follow-up email to this lead based on our previous thread, and show it to me before sending."

Notice the last prompt asks the agent to show you the draft first. That habit — review before send — is worth keeping.

Step 5: Keep agent access safe

Giving an assistant access to customer data deserves the same care as any integration. A few practices:

• Scope narrowly. Grant the least access the task needs. If a read-only key is enough, use it.
• Require approval for sensitive actions. Configure your CRM so that writes or outbound emails need human confirmation.
• Watch the audit log. Every action an agent takes should be recorded. Review it, especially early on.
• Revoke when done. If you connected an agent for a one-off project, disconnect it afterward.

Kantos supports all of these — OAuth 2.1 authorization, org-scoped access, rate limits, per-key audit logs, and optional human-in-the-loop approval. For the background on why this matters, see the CRM built for AI agents.

Troubleshooting

A few common snags:

• The agent says it has no tools. The server is not connected or authorization did not complete. Re-check the server URL and run the OAuth step again.
• Authorization fails. Make sure you are signing in to the correct CRM account and that your user has permission to authorize agents.
• The agent refuses an action. It may be hitting a scope limit or an approval gate — by design. Check what access you granted.

Where to go next

If you want the underlying concepts first, read what an MCP server for a CRM is. When you are ready to try it on your own records, join early access and connect your first agent in a few minutes.