mad-cli 0.4.0

Operator CLI for Mad — guided install, lifecycle, configuration, keys and version management for mad-edge containers.

mad-cli

Operator CLI for Mad. mad-cli installs, runs and manages mad-edge containers — the Mad runtime that provisions isolated workspaces and launches autonomous coding agents (Claude Code, OpenCode, …) against your repositories.

mad-cli is the thing an operator installs on a server or a Raspberry Pi. It does not run agents itself: it generates a Dockerfile, a compose.yml and a .env, builds the image and drives the container lifecycle through Docker Compose. The agent runtime lives in the mad-edge package, which is installed inside the container.

Alpha (0.x). Interfaces may change between minor releases. Pin a version in production.

Quickstart

pip install mad-cli

# Guided install: generates the config, builds the image, starts the container.
mad install

# Day-to-day lifecycle.
mad start        # build + up (detached)
mad status       # container state + API health
mad logs         # follow logs
mad stop         # stop, keeping workspace data on the host

mad install walks you through the instance name, host port, data path, GitHub token, git identity and Claude OAuth token, then writes everything under ~/.config/mad (override with MAD_CLI_CONFIG_DIR) and starts the container. Once it is healthy the Mad HTTP/MCP API is available on the port you chose.

Commands

The surface grows across the early minor releases; everything below the current line is planned.

Command	Since	What it does
mad install	v0.1	Guided setup: render config, build image, start the container.
mad start / stop / restart	v0.1	Container lifecycle (up --build / down / down+up).
mad status	v0.1	Container state and API health for an instance.
mad logs	v0.1	Follow container logs.
mad shell	v0.1	Open an interactive bash shell inside the container.
mad list	v0.2	List every configured instance.
mad info NAME	v0.2	Show one instance's resolved config.
mad keys	v0.2	Set, rotate and mask API keys / tokens in .env.
mad config	v0.2	Read and edit .env values safely.
mad versions / update	v0.3	Show the latest published mad-edge and rebuild onto it.
mad serve / service	v0.4	Run the local HTTP API, or install it as a background service.

Instance-scoped commands take an optional INSTANCE argument; when exactly one instance is configured it is used by default.

HTTP API (optional)

Every CLI capability is also exposed over a local HTTP API so a UI/dashboard can build on the same logic. It ships as an optional extra so the base CLI stays a two-dependency package (typer + rich):

pip install 'mad-cli[server]'

mad serve                 # foreground API on http://127.0.0.1:7373

The API binds 127.0.0.1 by default and requires a bearer token (auto-generated at ~/.config/mad/api-token, mode 0600) on every request except /health. Secret values are always masked on reads. OpenAPI is served at /openapi.json, docs at /docs. See docs/03-contracts/http-api.md.

To keep it running across reboots:

mad service install       # systemd user unit (Linux) / launchd LaunchAgent (macOS)

If the server extra is not installed, mad service install auto-provisions a dedicated virtualenv under ~/.config/mad/server-venv and installs the API there (use --wheel PATH to install from a local artifact instead of PyPI), then points the service at it — the base CLI never needs FastAPI in its own environment.

MVP limitation. Long operations (install with start, start, update) run synchronously — the request blocks until the Docker build and health wait finish. Background jobs are a future enhancement.

How it relates to mad-edge

	mad-cli (this package)	mad-edge
Runs on	the operator's host	inside the container
Console script	mad	mad-edge (server)
Role	install & lifecycle management	the agent runtime / HTTP + MCP API
Talks to	Docker, the filesystem, PyPI	Claude Code, OpenCode, GitHub, workspaces

mad-cli pins or tracks the mad-edge version through the generated Dockerfile; the container installs mad-edge from PyPI at build time. Version bumps follow the mad-edge versioning policy.

License

MIT.