How to Give Claude, GPT and Custom Agents Safe Write Access to Your CRM
Read access is easy to grant and rarely dangerous. Write access is where teams get burned. Here is a sequence for connecting an assistant built on Claude, a GPT-based workflow or a homegrown agent to your CRM, and how AI PRO CRM handles each step.
Step 1: register the agent as its own identity
Create an agent record with a name, a purpose and an owner. Do not reuse a human's login. The owner answers for the agent's behaviour. In AI PRO CRM the owner must be an active member, and disabling an agent stops its credentials immediately.
Step 2: choose the narrowest permissions
Most assistants need to read widely and write narrowly. AI PRO CRM has three presets. Analyst is read-only. Assistant adds one write: internal notes on companies. Operator is currently the same as Assistant. No preset lets an agent change deal stages, create tasks or edit fields.
Step 3: issue a dedicated credential
In AI PRO CRM each agent gets its own bearer credentials, shown once, stored only as a hash, expiring after 30, 90, 180 or 365 days, and individually rotatable and revocable. Never paste credentials into prompts; keep them in the agent's runtime configuration.
Step 4: decide what needs a human
A common pattern in governed systems is approval routing: low-risk writes go through, field changes are held for review, destructive actions are refused. That needs a governance layer, which AI PRO CRM does not have; external governance through WorQ is not connected yet. Instead, the agent can only add notes, and every change to a record is made by a person.
Step 5: run on test data first
Point the agent at a separate workspace with realistic but fake records, and read every note it writes before it touches real customer data.
Step 6: connect, then watch the audit trail
Review the agent's activity regularly. In AI PRO CRM, owners and admins can see an append-only audit trail of workspace changes, and every MCP call is recorded in a call ledger with the agent, credential, tool, status and record IDs. It never stores arguments, results or note bodies; the notes appear on the company with the agent as author.
Step 7: rely on limits as a safety net
AI PRO CRM applies a daily cap on successful writes per agent (100 on Free, 500 on Starter and Team, 2,000 on Business) plus rate limits of 120 MCP requests per minute per credential. A runaway loop or injected instruction hits the fence quickly. These are limits, not budgets.
Framework notes
- Claude and other MCP-capable clients: point them at
https://app.aiprocrm.com/mcp with the header Authorization: Bearer aiprocrm_ag_...; the agent's identity comes from the credential.
- GPT-based or custom agents: use any MCP client library in Python or TypeScript. Agents use MCP only; AI PRO CRM's REST API is for human users.
- Treat denied, rate-limited and cap-reached results as distinct outcomes, not errors to retry.
FAQ
Can I give one agent access to several workspaces?
No. In AI PRO CRM an agent belongs to one workspace, which keeps its permissions, limits and audit trail clean.
What if the agent needs to delete records?
It cannot. AI PRO CRM agents have no delete or update tools. An agent can add a note recommending an action for a person to take.