Claim: a fresh clone of mutx-dev/mutx-dev runs its control plane on SQLite with no account, enforces roles from the database, and returns block for a tool call that matches a tenant rule and allow for one that doesn’t.
Inputmain at 5ad48f0, a local SQLite file
NeedsGit, Python 3.11, uv, curl, sqlite3, no Docker, no API key, no account
Evidence/health JSON, CLI output, two policy decisions, 58 tests

Prerequisites

  • macOS or Linux with Git, curl and sqlite3;
  • Python 3.11 and uv (plain python3.11 -m venv and pip also work);
  • port 8000 free, and about 500 MB of disk for the Python packages.
The hosted service at mutx.dev and api.mutx.dev is down, so this recipe stays on localhost. Nothing in it calls a model.

1. Clone and install

This installs the API’s runtime dependencies, pytest, and the mutx CLI from the checkout into .venv.

2. Start the API on SQLite

Leave it running. SQLite needs WEB_CONCURRENCY=1. Turning off email verification lets a password account sign in without an email provider. The log ends with Database tables verified/created and Starting background agent monitor.... Open a second terminal in the same mutx-dev directory and check the probe:

3. Give the CLI a throwaway config

The CLI keeps its API URL and tokens in ~/.mutx/config.json. Point HOME at a temporary directory in this terminal so your real config isn’t touched:
The CLI’s default URL is http://localhost:8000, so no extra flag is needed here.

4. Create and sign in an operator

register prints “Verify your email” even with verification off. The password needs an uppercase letter and a special character. Keep -u on login, because the login command treats localhost:8000 as “not set” and otherwise sends your password to https://api.mutx.dev (resolve_hosted_api_url).

5. Hit the role check, then grant the role

This fails, because every new account is a VIEWER and creating agents needs DEVELOPER. No API route grants roles, so write it into the database:
Roles are read from the database on each request, so you don’t need to sign in again.

6. Register and stop an agent

Creating the agent stores a record with status creating and doesn’t contact OpenAI. Stopping it sets the status to stopped in the database.

7. Add a block rule and evaluate two tool calls

The CLI has no policy command, so use the API with the token the CLI saved:
policy_store.py matches each rule’s pattern against the tool name as a glob or a substring, case-insensitively.

8. Run the matching tests

These cover policy CRUD and evaluation, approval enforcement before tool execution, role checks and run records.

Expected evidence

  • /health returns "status":"healthy" and "database":"ready";
  • the first mutx doctor shows Authenticated: no and API Health: healthy, with OpenClaw missing and documents not ready;
  • mutx whoami prints Email: operator@example.com and Plan: FREE;
  • the first agents create prints Error: Insufficient permissions. Required roles: ['DEVELOPER'], and the second prints Created agent: <uuid> - demo-agent;
  • agents status ends with Status: stopped;
  • the shell.exec evaluation returns "decision":"block" with one match from no-shell, and the web.search evaluation returns "decision":"allow" with "No enabled policy rules matched";
  • pytest prints 58 passed.
These matched on 23 October 2026 at 5ad48f0.

This recipe does not prove

  • that an agent ran. The agent is a database record, and no process registered against it;
  • that stopping it would stop a live agent process. The API writes no command to the queue agents poll;
  • that the governed runtime in src/security, its signed receipts or the approval flow ran. /v1/policies/evaluate is the tenant rule check only, and the approval route needs a paid plan;
  • anything about PostgreSQL, Redis, Docker, the dashboard, the macOS app or the hosted service;
  • that this setup is safe to expose. Email verification is off, the JWT secret is generated per process, and the role was granted by editing the database.
Stop the API with Ctrl-C, then remove the clone and the temporary config from the second terminal: