Five-minute quickstart¶
This workflow uses the built-in mock provider. It sends no request to a paid model API.
1. Create local configuration¶
From the project you want to protect:
This creates:
agent-loop-guard.yml
.agent-loop-guard/
└── profiles/
├── claude.env
├── cline.txt
├── codex.toml
└── opencode.json
The command generates a gateway key when it creates a new configuration. Re-running setup preserves the existing key unless --force is used.
2. Check the environment¶
Docker may show WARN. That does not prevent Guard, MCP, Replay, or Benchmark from running.
3. Start the local gateway¶
Keep this terminal open. The default addresses are:
| Surface | URL |
|---|---|
| Dashboard | http://127.0.0.1:8787 |
| OpenAI-compatible API | http://127.0.0.1:8787/v1 |
| Anthropic-compatible API | http://127.0.0.1:8787 |
| API documentation | http://127.0.0.1:8787/docs |
4. Run a guarded demo¶
In another terminal:
The command creates a session with repeated requests. In the default Shadow mode the requests are recorded and flagged without being blocked.
Open the dashboard:
Inspect:
- Sessions for request counters and loop decisions;
- Replay for spans, events, timing, and token usage;
- MCP Firewall for configured servers and pending approvals.
5. Try Enforce mode¶
In Enforce mode a matching blocking policy returns a local error instead of forwarding the request.