Fields agents never see
Mark a field private and it never leaves the server over MCP. Agents cannot read, search or write it, whatever model they run. Only a person can change that.
For teams whose people already work with AI agents
Every employee brings their own Claude Code, Codex or Cursor. Each one gets its own token and its own rules. Every change an agent makes is previewed, approved when it has to be, and can be undone with one click.
of full-time workers in a Clutch survey have built their own AI agent.
Clutch, Sept 2026of those employee-built agents can reach company data, customer records included.
Clutch, Sept 2026of employees admit using AI tools without approval, often from personal accounts.
Okta, AI Agents at Work 2026How it works
Who you sell to, your industry, how customers find you, what you track. Click cards, pick from lists, or type your own when nothing fits. Your answers become objects, fields and pipeline stages for your business.

Maja's Cursor, Lars's ChatGPT connector, your Claude Code: each gets a token and a rule set. Read only, edit records alone, or ask before touching the structure. Revoke one token without breaking the others.

When an agent wants a new field or a new view, it lands in Changes with a line-by-line preview. Nothing is written until you approve. Anything applied, by a person or an agent, can be reverted.

Control
Mark a field private and it never leaves the server over MCP. Agents cannot read, search or write it, whatever model they run. Only a person can change that.
Each search, read and proposal is in the activity log with the agent's name. You see what your agents look at, not only what they change.
Every change stores its inverse. Revert a whole import, a renamed stage or a deleted record, and see if someone changed it since.
Data must stay in the EU and someone uses a model hosted outside it? Email and phone are hidden from agents by default, and the setup tells you why.


Your AI
Other CRMs sell you their assistant and bill you per credit. Here your team keeps the tools they already pay for. The CRM never calls a model itself, so there is nothing to bill.
claude mcp add --transport http agent-crm \
https://your-workspace/mcp \
--header "Authorization: Bearer acrm_…"
The Agents page gives you this line with a real token, plus the same for Codex and Cursor.
Daily use
Tables and a pipeline built from your own data model. Light or dark, six accent colors, three densities. Drag the handle.

The first test
Agent CRM is in its first phase. It works, it is tested, and it has not met enough real teams yet. That is what this test is for.
FAQ
No. The CRM never calls a model. Your agents call the CRM over MCP, with your model and your tokens. Fields you mark private never leave the server.
During the test we agree the setup with each team. The plan for general use is hosting in the EU. If your data has to stay in one country, tell us in the form.
Nothing during the test. After it, a flat price per workspace. AI is never billed, because you bring your own.
Anything that speaks the Model Context Protocol over HTTP: Claude Code and Desktop, OpenAI Codex, Cursor, Gemini CLI, local models and your own agents. Each one gets its own token.
Its change shows up in Changes with a preview. Changes to the structure and all deletes wait for approval by default. Anything applied can be reverted, and you can revoke the agent's token in one click.
Today your agent can move records in through a reviewed change. A one-click CSV import is next on the roadmap.