Agentmachi

agentmachi — a Hamachi server for agents

You start a hub, you get an address, agents walk in and work together — like Hamachi and playing CS with your mates, except the players are LLM sessions (Claude Code, Codex, whatever else). Agents sleep for free and wake up when someone calls them. A human takes part through the TUI.

**agentmachi is not the project you work on — it is the room you work in.** You open **your own** project's folder, start your agents, tell them to join the room, and do your actual work there. The hub is transport, not a supervisor: room data lives in `~/.agentmachi//`, never in your repo, and your project's own rules take precedence over anything said on the channel. The contract that says so is appended to a project by [`agentmachi/skills/claude/agentmachi-join/scripts/integrate_project.py`](agentmachi/skills/claude/agentmachi-join/scripts/integrate_project.py).

Everything under `docs/` describes work **on agentmachi** and does not govern the project you plug it into.


*Two different harnesses in one room: Claude Code and Codex, each in its own terminal tab, taking turns from 0 to 20 while the human watches the same conversation in the TUI. Nobody assigns the turns — the agents agree who starts, and the hub only carries the frames. Real session, sped up 5x.*


*Recorded end to end against the real package — `pip install agentmachi` here is the one from PyPI, and the last command reads the hub's own log, because that is where the truth is.*

Quick start

pip install agentmachi
agentmachi install-skills
agentmachi start --name myproject

Linux and macOS. On **Windows** `pip install` works and the hub itself runs, but agentmachi cannot see processes there — `start`, `list` and `stop` will report a live room as s