Put policy before
an agent action.

Local policy contracts

The TypeScript line contains strict v1 contracts for requests, decisions, permits, outcomes, and evidence. They are independently versioned from the Python compatibility package.

Compatibility is not a firewall

The shipped OpenClaw hook and MCP HTTP/in-process adapters can observe policy decisions and return a block result. They can also be disabled or bypassed, so they are telemetry, not capability separation.

Proof before stronger claims

A same-UID configured intercept exists for one GitHub file write. Capability isolation and firewall results are not claimed. See the named-action proof.

Choose a package

The TypeScript and Python packages have independent release lines and different assurance scope.

pip install agentguard-tech

Evaluate a proposal

The TypeScript package can load a YAML policy and evaluate it in-process — no network call, no API key.

import { PolicyEngine } from "@the-bot-club/agentguard";

const engine = new PolicyEngine();
engine.loadFromYaml(yaml);
engine.evaluate(request, ctx, policy.id);

Keep deferred states blocked

V1 can allow or block. A require-approval decision blocks dispatch; human approval is deferred.

id: agent-baseline
name: Agent baseline policy
version: 1.0.0
default: block
rules:
  - { id: allow-reads, action: allow, when: [{ tool: { in: [read_file, search] } }] }
  - { id: approve-email, action: require_approval, when: [{ tool: { in: [send_email] } }] }
  - { id: block-shell, action: block, when: [{ tool: { in: [shell_exec] } }] }

Read the limitations

The docs page documents which adapters are compatibility-only and which broker proof is still missing.

agentguard.tech/docs/#limitations