colonizer-harness

The mothership of Colonizer: the colonizer command.

Every task gets a colony: its own KVM microVM with a fresh git worktree and a coding agent inside. The agent can do anything in there. The colony is linked back to your machine, the mothership, so its chat and a real terminal are one hop away. When the agent needs you, it asks with choices. When the work is done, the mothership commits it and opens the pull request.

Install

curl -fsSL https://colonizer.dev/install.sh | sh

This crate is the harness's Rust source. It isn't an install on its own: cargo install colonizer-harness builds only the colonizer binary, and the app also needs microsandbox, the in-VM daemon, the agent module and the web UI beside it, which the installer puts there. The install guide covers the installer and building from source. Linux x86_64 with KVM, or an Apple Silicon Mac.

What it does

• A microVM per task. Not a shared-kernel container: a colony that goes rogue can wreck its own worktree, and that is all.
• A private mesh home. Colonies join a private network with the mothership, never your own tailnet, and can't reach each other. On a Mac, colonies are reached on a loopback port instead.
• Choices, not walls of text. Questions arrive as cards with an "Other…" answer, in a web UI that shows each colony's chat and terminal side by side.
• Placeholders only. The GitHub token never enters a colony, and the agent's API credential is swapped in by the sandbox's host-side TLS proxy, for one host, on the way out.
• The host publishes. The microVM is gone before the mothership commits, pushes and opens the pull request.

The crates

Documentation

The install guide, the vision, the architecture and the protocol are on colonizer.dev, mirrored from the repository's docs/. Code, issues and releases: Colonizer-dev/harness.

MIT.