AI agent governance API and MCP documentation

Maetra is a runtime control plane for AI agents. Use the REST API or MCP server to keep work aligned with the user's task, pause sensitive actions for policy or human approval, scan risky content, and preserve evidence of what happened after approval.

New to Maetra? Start with the Quickstart. Connecting an agent or assistant? Use the MCP server guide. Just need an endpoint? Jump to the API reference.

The three runtime controls#

Task Guard

AI agent task alignment. Turn the user's objective, constraints, scope, and success criteria into an active task contract. Before a material action, Task Guard checks whether the work is aligned, supporting, needs an explanation, requires user confirmation, or should be refocused. Afterward, it can compare expected and reported effects.

Govern

Human-in-the-loop approvals and execution receipts. Before an agent moves money, deletes data, or emails a customer, it calls Maetra. Exact policies or decision intelligence can approve, block, or route the action to a human. A signed decision is bound to the exact action payload, and execution receipts can link that approval to attempts and the observed result.

Secure

Runtime content scanning. Screen prompts, tool calls, and model outputs against your workspace rules — prompt injection, data exfiltration, disallowed tools — and get back a verdict (safe, flagged, blocked) and a recommended action. Anything flagged or blocked becomes an incident.

Two ways to integrate#

Both are backed by the same evaluation engine and authenticated with the same workspace API key.

What you can build#

• Payment and finance agents that cannot change the approved account, amount, or currency after review.
• Customer operations agents that pause before refunds, account changes, or external messages.
• Long-running assistants that re-check their task contract before material actions and report actual effects afterward.
• MCP-connected agents that discover only the Task Guard, Govern, and Secure tools enabled for their workspace and API key.

Approval and execution are separate facts. A decision token proves what Maetra authorised; an execution receipt records how that exact action was attempted and what effect evidence was observed.

Good to know You don't need to register your agents first. Every endpoint accepts a free-text agent_name, so you can integrate in minutes and register agents later for richer reporting. Govern still evaluates API-only and MCP agents against organization-wide policies; when agent_name or agent_id matches a registered agent, agent-scoped policies can apply. See Agents.

Base URL#

All REST endpoints are versioned under /v1. The MCP server has its own base URL — see the MCP overview.

Next steps#

1. Create an API key in the dashboard.
2. Run your first checkpoint and scan, connect an MCP client, or start a Task Guard task.
3. Browse the full API reference.