Questions, answered.

A lane is one project on one machine. It reads your backlog in order, takes the next item it can work, and runs that item to the end — in its own AI session, on its own branch and worktree — before it takes another. A lane works one item at a time, so more lanes mean more items at once.

Claude Code, the reference, with the farm’s full feature set; OpenAI’s Codex CLI; Google’s Gemini CLI; and OpenCode, which reaches Grok, Mistral, DeepSeek, OpenRouter or a model on your own GPU. You choose per machine, per project, even per item, and switch from the panel. Aider, Goose and Qwen Code are coming.

A machine — a host — is a computer you already own with the farm installed on it: a Linux box for most of the work, a Mac for iOS and macOS builds because it is the only machine that can sign them, a Windows machine when you ship Windows software. Your code, your git server and your tests stay on it, and nothing on the build path goes through us.

Reach, and nothing else. Hosts, lanes, sessions, tests and the commits on your main branch all work with the hub switched off. With it, the app and any browser get to your farm from outside your network, and what the farm builds can be deployed for you. It comes with the hub tiers; a farm without it is still a whole farm.

The farm goes idle cleanly. No new project, item or session starts, and work in progress is stopped properly — the lanes just stop. Nothing is deleted: your repositories, backlogs and history stay exactly where they are, on your machines. Renew, and it picks up where it left off.

You do, straight to your AI provider — Anthropic, OpenAI, Google or others — or nothing at all for a local model. Bring your own subscription or API key: we never resell tokens and no tier marks them up. The pricing page says what to budget per lane per month. What you pay us is the licence for the farm itself.

Your tokens are your money, so the farm spends them carefully. Routine items run on a lighter model and step up only when a retry needs it. Every session has a context and turn budget. A small scout maps the code before the main session reads anything, and big files and images are handled in short sessions of their own. Nothing sits polling while it waits. And a usage governor pauses the lanes before your weekly allowance runs out.

On your machines, and only there. Your repositories, your git server, the worktrees, the test runs and the build output all sit on the hosts you installed the farm on. If you take the hub, it carries the app's requests to your farm; it is not a place your code is copied to.

No. The AI session edits and commits in a working copy of its own, and that is all. Your machine then re-runs your tests, rebases onto main and pushes, with a git token kept on the machine that the session is denied reading — as it is denied ~/.ssh, .netrc, .env files and signing keys. Only what passes lands.