git-mp-agent — the AI orchestrator

Turns a Discord forum into an AI work queue. Open a thread, tag it with a repo, and git-mp-agent spins up a sandboxed coding agent in a container — with a hard per-thread budget on LLM spend.

1. Run

git-mp-agent serve --port 9100
# one process: Discord bot + metering LLM proxy

2. Configure the agent

All settings live in the (encrypted) config table. On first run the global row is seeded with REPLACE_ME placeholders — the server comes up, the bot/proxy stay idle — then set discord_tok, llm_upstream and llm_api_key at /config.

3. Map channels to repos

Each watched forum channel gets a config entry: the repos it may work on, the container image to run, and default_limit_usd — the spend cap per thread.

4. How a task runs

1. teammate opens a forum thread, picks a repo tag
2. agent starts a container with a one-off LLM token
3. every LLM call is metered through the proxy
4. budget crossed → container stopped, thread notified
5. results land as ordinary git pushes + thread replies

Because the proxy injects the real API key server-side, the agent container never sees your credentials — only an opaque, budget-limited token.