MCP CRM Integration: A Practical Guide to Connecting Agents to Customer Data
The Model Context Protocol (MCP) has become a common language between AI assistants and the tools they use. For a CRM that matters: instead of every agent framework needing a bespoke connector, one MCP server can expose companies, people, deals and tasks to any MCP-capable client.
This guide explains how a CRM should expose itself over MCP, using AI PRO CRM as an example.
What an MCP server for a CRM looks like
An MCP server publishes tools with typed inputs and outputs. AI PRO CRM's remote endpoint, https://app.aiprocrm.com/mcp over Streamable HTTP, publishes ten read tools: workspace_summary, crm_search, companies_list, company_get, people_list, person_get, opportunities_list, opportunity_get, tasks_list and activity_timeline. It publishes one write tool, activity_note_create, which adds an internal plain-text note to an existing company.
Reads are broad, because an agent without context gives poor answers. Writes are deliberately narrow: an agent can record what it found, but deal stages, tasks and fields stay with people.
Some CRMs let agents update fields. If you go that way, look for a governed write path where risky changes are held for a human to approve. AI PRO CRM does not have that layer yet, which is why its agent writes are limited to notes.
Authentication
Each MCP connection should authenticate as a specific agent, not as a user. In AI PRO CRM an agent is a named identity with a handle, a purpose, an owner and a status; only active agents can authenticate.
The agent sends Authorization: Bearer aiprocrm_ag_.... Credentials are issued per agent, shown once, stored only as a hash and expire after 30, 90, 180 or 365 days. You can rotate or revoke one immediately, and disabling the agent stops its credentials at once without affecting anyone else.
Matching tools to the agent's job
AI PRO CRM uses presets rather than fine-grained scopes. An Analyst agent gets the read tools only. An Assistant gets the read tools plus activity_note_create. An Operator currently has the same tools as an Assistant.
Handling write outcomes
Treat each result explicitly.
- Success: continue.
- Denied: the preset does not allow the call. Report it; do not retry.
- Rate limited: back off. The limit is 120 MCP requests per minute per credential and 600 per workspace.
- Daily write cap reached: stop writing for the day. Each agent has a cap on successful writes (100 on Free, 500 on Starter and Team, 2,000 on Business) as a safety fence against runaway loops.
Common mistakes
- Exposing raw record IDs as the only lookup. Give agents a search tool and let them resolve IDs from the results.
- Treating logging as optional. AI PRO CRM's call ledger records the agent, credential, tool, status, duration and record IDs for every call, but never arguments, results or note bodies.
- Skipping idempotency.
activity_note_create requires an idempotency key: a replay returns the original note and a changed payload with the same key is rejected, so retries never duplicate notes.
FAQ
Does MCP replace the REST API?
No. MCP suits interactive agents; REST suits programs written by people. In AI PRO CRM, agents use MCP and the REST API is for human users with personal tokens.
Which clients work with an MCP CRM?
Any MCP-capable client that supports remote servers over Streamable HTTP with bearer authentication.
Can an AI PRO CRM agent update deals?
No. Agents can read CRM records and add internal notes to companies. Changes to deals and tasks are made by people.