How it works

From first agent to enforced policy in four steps.

Put the gateway in front of your MCP servers, give every agent an owner and a scope, watch what would be blocked, then enforce. The same steps work from the console, the REST API, the TypeScript SDK and the GitHub Action.

  1. 01

    Put the gateway in front of your MCP servers

    Register each MCP server your agents use. Agent calls go through AtmanID, which checks them before forwarding.

  2. 02

    Give every agent an owner and a scope

    Register agents in the console, through the API or from a GitOps manifest. Each gets a short-lived token for the tools it may call, and a person accountable for it.

  3. 03

    Watch in shadow mode

    New servers start in shadow: every call is checked and logged, and what would have been blocked shows as a would-deny — nothing breaks.

  4. 04

    Enforce, and prove it

    Switch a server to enforce after a preview of what it will block. Every allow and deny lands on the tamper-evident audit trail, ready for an evidence pack.

Iris · natural-language governanceBounded by your own role

you  > grant me admin on github-prod

iris > Denied. Your role (End User) does not hold GRANT on github-prod.

• Same policy decision point as POST /api/v1/* — I can’t exceed your access

• Logged: POLICY_DENY · audit ref req_c41a…

• Opened REQ-0231 instead — approver dev.torres (owner) · SoD check clean

Denials are evidence too — every one is queryable.

Illustrative transcript — mirrors the live Iris response shape. See what Iris does
GitHub MCP server · shadow mode · last 7 daysBefore you enforce

1,284 calls checked

Would have been blocked

  • Tool outside the agent’s scope9
  • Agent over its rate limit3

Agents affected: Deploy Bot, Release Notes Agent. These calls went through while in shadow mode.

Next: switch to enforce. Every decision is recorded on the audit trail.

Illustrative. Prefer code? The same steps through the API are on the developer portal — no login required.