How it actually works
Guard is an MCP (Model Context Protocol) integration. It sits in front of an agent's MCP tool calls and checks each one against a sealed .nomos artifact before the call is allowed to execute — not a generic policy engine, a specific enforcement point in the MCP tool-call path.
npx nomos-guardPut a sealed policy in front of an MCP tool call in under a minute. No account, no setup.
Two ways to govern
Path 1 — Direct integration
Your agent calls NOMOS explicitly, through three tools: verify a decision against a sealed artifact, turn a policy document into one, or audit a policy for contradictions before you seal it.
verify_decisiongenerate_governancedetect_contradictionsPath 2 — NOMOS Guard (the gateway)
Wrap an MCP server you already run — GitHub, Postgres, Stripe, any of them — with zero code changes to the server itself. Guard sits in the execution path and checks every tool call before it reaches the real system.
- GitHub MCP — block force-pushes to main, gate repo deletion
- Postgres MCP — deny destructive queries against production
- Stripe MCP — escalate refunds above your approval threshold
npx nomos-guard \
--artifact pub_database_policy \
-- postgres-mcp-serverAn unrestricted agent, governed
Agent:Clean up the old customer records — drop the table.
DROP TABLE customers;
checking pub_database_policy…
DENIED by governance policy
Rule R1 — Production data destruction is prohibited without dual approval.
Audit hash: sha256:8f3a2c1e9d04b7f6…
Agent:Understood — I can't perform that action. Escalating to your team instead.
Illustrative — MCP servers run over stdio, not a browser session, so this is a recorded transcript rather than a live call.