mad-bros 0.4.0

Multi Agent Develop — self-hosted autonomous Claude sessions against GitHub repos.

Mad About

That's mad!

Multi Agent Develop — a multi-agent system designed to build software autonomously. It takes an idea and drives it end-to-end: from the first line of code to a working product.

What is this?

Mad About orchestrates a team of AI agents that collaborate to design, implement, test, and ship software without a human in the loop for every step. You give it a goal; it figures out the rest.

Status

Early days. The first milestone is Mad — a self-hosted API that provisions workspaces, clones repos, and runs Claude agents autonomously against them.

Install

Mad ships as a pip-installable Python package (mad). From a checkout:

make install   # create venv + editable install with dev deps
make test      # run the pytest suite
make serve     # uvicorn factory (override HOST=/PORT= if needed)
make help      # list every target

All commands are wrapped by the Makefile; the raw equivalents live in pyproject.toml (the mad console script) and in the project documentation.

Project structure

mad/
├── pyproject.toml          # package metadata, deps, `mad` console script
├── src/mad/
│   ├── api/                # FastAPI app + routes (thin HTTP layer)
│   │   ├── app.py          # create_app(store=...) factory
│   │   └── routes/         # sessions, events, stream
│   ├── core/               # domain — session log, workspace, security, SessionStore
│   ├── agent/              # harness loop + tool execution
│   ├── providers/          # LLMProvider protocol + claude_cli / anthropic_api / fake
│   └── cli.py              # `mad` console entry-point
└── tests/                  # pytest acceptance + security tests

Hard rules and conventions that govern every change live in CLAUDE.md.

Documentation

• docs/backlog.md — known improvements deferred past v0.1.
• docs/sandbox-bwrap.md — hardening guide for the execution sandbox using bubblewrap.

License

See LICENSE.