Coding agents settle your GitHub issues, each in its own KVM microVM, and come back with a pull request. Open source, MIT, on your own machine.
Developer toolsRust · TypeScriptGitHub @Colonizer-devOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
The open-source core of Colonizer: coding agents in private KVM microVMs, linked to your machine by a private mesh, asking with choices and returning pull requests. MIT.
Taken from the repository README
Loading the full README. The saved excerpt is shown below.
Coding agents are good enough to work on their own for a long time. Most setups still make you pick
between safe (a sandbox, one task at a time, an approval every few seconds) and fast (an agent
with your credentials loose on your laptop). This is the setup that doesn't make you pick.Read the full README ↗cargo install colonizer-harness --locked && colonizer setup
# no Rust toolchain? curl -fsSL https://colonizer.dev/install.sh | sh
Not ready for unattended work on sensitive repositories. An external audit of v0.1.3 found four ways past the wall that keeps secrets on the host, and they are not fixed yet (docs/audit.md).
You need git, gh, curl and tar. On Linux — or Windows 11 through
WSL2 — give your user /dev/kvm
(sudo usermod -aG kvm "$USER", then log out and back in) and install native Claude Code, which
colonies run. Install the latest release with a
Rust toolchain (1.88 or newer), or without one by piping the installer script instead — the crate is
on crates.io as colonizer-harness, and setup fetches that release's own installer for the version
it installed. Either way the release is checked against its SHA256SUMS and its build attestation,
and ~/.local/bin/colonizer is linked (if ~/.local/bin is not on your PATH, run that path); then:
colonizer # starts the mothership, prints a sign-in link and opens it
colonizer open # reprints the link
colonizer --help # the rest
To build from source (Node.js 20.19+ or 22.12+, and Rust 1.98+):
git clone https://github.com/Colonizer-dev/harness && cd harness
scripts/install.sh # builds everything into ./dist
dist/bin/colonizer
The colonizer binary on your machine is the mothership: it runs the web UI, a bundled Headscale and
userspace tailscaled, and the provider gateway. Each colony is a microVM running colonizer-agentd,
which supervises the agent runner (Claude Code by default) against /workspace. The agent contract is
JSON Lines on stdio, so an agent module can be written in anything. Colony output is untrusted until
the mothership has sanitized it and opened the pull request. The diagram is in
docs/vision.md; the design is in docs/architecture.md.
| Install · Configuration · Updates · Releases | Requirements, first run, every setting, budgets and quotas, upgrading |
| CLI · MCP · Colonies · Fleet | Driving a mothership, colony lifecycle, pairing motherships |
| Architecture · Protocol · Conformance | Modules, repository layout, the wire format, UHP conformance |
| Trust model · Boundaries · · |
Try the cockpit without a mothership: (cd web && npm run dev), then open
http://127.0.0.1:5173/?mock=1 for the UI against an in-browser mock backend. Every test suite and
script is in docs/development.md.
See CONTRIBUTING.md and SECURITY.md. MIT, see LICENSE. Vendor logos in the UI are CC0 artwork from Simple Icons; the marks stay their owners' trademarks. See NOTICE.
| Where secrets live, what is enforced, what the audit found |
| Providers · Loops · Jev · Hosted | Model providers, built-in loops, second opinions, the hosted contract |
| Bench · Vision · Gaps · Decisions | The fixed task set, what it measures and what it does not, the roadmap, what was decided against |
| CHANGELOG · changelog.d/ | Every release, and what has merged since |
| Development · Runner authoring · Good first issues | Tests, a new agent module, starter issues |