A self-hosted infrastructure layer for running coding agents. mad provisions an isolated workspace, clones a GitHub repository into it, and launches an external coding agent against the code — then streams every line of the agent's stdout to you as agent.output Server-Sent Events, per session.

• Isolated by design — each session gets its own workspace, its own agent process, and its own event stream.
• Runs many at once — multiple parallel sessions, no shared state between them.
• Bring your own agent — Claude Code CLI today, plus an opencode provider.
• Infrastructure, not an orchestrator — mad launches agents and reports what they do. It does not parse tool calls, execute tools, or manage a conversation loop. That belongs to the agent's own harness.

Status: early days — 0.x. The HTTP + SSE surface is stable enough to build clients against; scope is deliberately small and honest about it.

pip install mad-edge     # distribution: mad-edge · console script: mad-edge
mad-edge serve           # HTTP + SSE API on 0.0.0.0:8000

Prefer a guided, containerized setup? mad-cli gives you the mad command: pip install mad-cli, then mad install walks you through provisioning a mad-edge container end to end.

Then create a session (provisions a workspace, clones a repo), send it a prompt, and stream the results:

curl -N http://localhost:8000/v1/events/stream   # live agent.output events

mad-core is AI-first: the day-to-day writing and maintenance of these repositories is done by AI coding agents (Claude Code), working from issues and conventions checked into each repo. Humans set the direction, define the guardrails, and review every change before it merges. The tooling below exists to make that model repeatable and safe.