MUTX is a FastAPI service that stores agents, deployments, runs, policies and approvals per user, with a Next.js dashboard, an Electron macOS app, a Click CLI and a Python SDK in front of it. It’s for teams that run agents in their own processes and want one place to register them, record what they did and check tool calls against rules before they execute.
StatusShipped, v1.4.0 source, hosted service offline
RuntimePython 3.11, FastAPI, SQLAlchemy on PostgreSQL or SQLite, Next.js 16, Electron
Try itLocal API on SQLite, no account, recipe
Proof hooks2,897 collected pytest tests, ruff, AgentDojo policy replay

How it works

src/api/main.py builds the app. Importing it at 5ad48f0 gives 251 routes: 246 under /v1 and the root probes /, /health, /ready, /metrics and /release. On startup it creates tables, starts a background agent monitor and a scheduler loop, and leaves the document and reasoning workers off unless configured. SQLite works with one process (WEB_CONCURRENCY=1); src/api/config.py rejects it in production. Routes that read or change a user’s data check roles through require_roles in src/api/auth/dependencies.py. Roles live in the database. User.roles defaults to ["VIEWER"], and signup, SSO and the localhost-only /v1/auth/local-bootstrap all create viewers. Creating agents or policies needs DEVELOPER, and reading audit events needs ADMIN or AUDIT_ADMIN. No route grants a role, so on a fresh install someone has to write the role into the users table. Agents don’t run inside the API. In src/api/routes/agents.py, POST /v1/agents/{id}/deploy writes a Deployment row and sets the agent to running, and POST /v1/agents/{id}/stop marks the agent and its deployments stopped. The agent itself is a separate process. sdk/mutx/agent_runtime.py registers it, sends heartbeats, metrics and logs, and polls GET /v1/agents/commands (src/api/routes/agent_runtime.py). Runs are records a client posts after the fact. There are two stores: POST /v1/runs with its traces (src/api/routes/runs.py), and /v1/observability/runs, which the SDK’s observability.py and mutx observability runs use. A run posted to one doesn’t show up in the other. Policy is checked in two places. Tenant policies in src/api/services/policy_store.py are named rule lists. Each rule has a type (block, allow or warn), a glob or substring pattern, and a scope (input, output or tool). POST /v1/policies/evaluate matches every enabled rule and returns block if any block rule matched, then require_approval, then warn, else allow. The governed runtime in src/security/ has its own PolicyEngine with allow, deny, modify and defer, signs each decision into an Ed25519 receipt and links receipts into a per-session hash chain (docs/governed-runtime-spec.md). POST /v1/security/actions/evaluate runs that engine as a dry run and stores the signed receipt. The two meet in ToolExecutionHandler.execute_tool in src/api/services/agent_runtime.py. It asks the governed runtime first, then the tenant policies through evaluate_with_approval (src/api/services/policy_approval.py). A deferred call runs only after it consumes a matching approved request, and the handler blocks the call if the approval lookup fails. This handler serves LangChain agents built by AgentRuntime.create_agent. No route in src/api/routes/ calls AgentRuntime at 5ad48f0, so the only way to exercise this path is to embed the runtime in your own code. The CLI’s mutx setup local reuses an API already answering on localhost:8000. Otherwise it runs scripts/dev.sh up with Docker, from the current checkout or from a managed checkout under ~/.mutx/runtime/local-control (cli/local_control_plane.py). Either way it then installs or imports OpenClaw, the only provider it accepts (cli/commands/setup.py). mutx doctor (cli/commands/doctor.py) prints the API URL, auth state, API health and the OpenClaw and document-worker readiness.

Policy replay on AgentDojo

benchmarks/governed-runtime replays recorded AgentDojo runs through PolicyEngine and the audit chain. The data is AgentDojo’s run logs (MIT licence) at commit 089ed468, benchmark v1.2.2, for gpt-4o-2024-05-13, claude-3-5-sonnet-20241022, claude-3-7-sonnet-20250219 and gemini-1.5-pro-002, with no attack and with the important_instructions attack. The policy denies delete_*, cancel_* and two account tools, allows get_*, read_*, search_*, list_* and check_*, and defers everything else. From results.md at d5093fe: The policy reads tool names only. A deferral protects only if the approver refuses, and the same rule that holds an injected send_money holds every legitimate one, which is why half the benign tasks needed approval. Most misses are one Slack task that sends the agent to an attacker URL through the read-only get_webpage. The traces are recorded, so a denial can’t show what the agent would have done next. The run covers one attack and four models, and only the policy engine and the chain took part. The API and dashboard weren’t involved. The README’s latency figures (p50 4.1 us for evaluate) differ from those in results.md (p50 4.2 us); the decision counts match. The replay itself wasn’t rerun for this page because it needs a sparse clone of AgentDojo. Its unit tests pass (see below).

Verify it

From a fresh clone at 5ad48f0 with Python 3.11 and uv, on 23 October 2026:
Ruff printed All checks passed!, collection printed 2897 tests collected, the API subset gave 58 passed and the audit-chain and benchmark tests gave 67 passed. The CI test command, pytest tests/api tests/unit/python tests/test_*.py, was stopped at a 330-second limit after 1,900 of its 2,867 tests, with 2 failures. test_ci_compose_release_gate.py::test_frontend_ci_builds_and_smokes_the_standalone_production_artifact fails because the release browser smoke no longer checks data-boundary-surface="dashboard". test_autonomy_work_order_sandbox.py::test_assessment_rejects_git_hook_tampering writes a hook into .git/hooks and fails on a machine with a global core.hooksPath, as this one has. To count the mounted routes:
It printed 251 246. The frontend checks (npm run build, npm run typecheck, npm run test:app, npx playwright test) need npm install and weren’t run.

Limits

  • The hosted service is down. On 23 October 2026 mutx.dev and mutx.dev/download/macos returned 503, app.mutx.dev returned 404, and api.mutx.dev, the CLI’s default API, answered “Application not found”. docs.mutx.dev returned 200. mutx setup hosted and the README’s dashboard link can’t work until that changes.
  • The README’s install paths are out of date. pip install mutx finds nothing because neither mutx nor mutx-cli is on PyPI. The published Homebrew formula installs CLI 1.2.0 from a March 2026 commit and declares an MIT licence, while the copy in this repo points at tag cli-v1.2.0 and declares BUSL-1.1. pyproject.toml on main says 1.4.0.
  • The v1.4.0 release (9 April 2026) has no assets. The newest macOS DMGs are on v1.3.0.
  • CI is switched off. The CI and CI Health workflows are disabled in the repo settings. The last CI run on main, on 21 September 2026, failed, and CodeQL failed on the 23 October push.
  • Stopping an agent changes database rows only. No code in src/api creates the Command rows that /v1/agents/commands serves, so an agent process isn’t told to stop by the API.
  • Budgets are read-only. The API mounts GET /v1/budgets and /v1/budgets/usage, and mutx budgets create, update and delete print that those routes aren’t mounted.
  • POST /v1/policies/evaluate-and-request-approval also requires a paid plan.
  • Runtime governance through Faramesh needs the separate Faramesh daemon on a local socket (cli/faramesh_runtime.py). mutx setup needs OpenClaw. Document workflows need predict-rlm, Deno and an OpenAI key.
  • The platform is licensed under BUSL-1.1. Internal production use is allowed, offering it to third parties as a hosted or embedded service isn’t, and each version becomes Apache-2.0 36 months after release. The SDK is Apache-2.0. GitHub reports the repo licence as NOASSERTION.

Source

Run the API locally and check a policy

Start the control plane on SQLite, sign in with the CLI, register an agent and watch a block rule fire.