start as many as you need. each one works in its own copy of your repo and comes back with a pull request. it keeps what it learns from you, so later conversations need less explaining.
needs git and node 22.18 or newerhow code gets written, tested and merged without you babysitting it.
each conversation works in its own git worktree, on its own branch. node_modules, builds and .env come along as copy-on-write clones, so it starts straight away.
run as many as you like. you keep coding in your folder the whole time, and nothing lands there until you merge.
a terminal next to the conversation that it can read. chrome to open your app, click through it and take screenshots. desktop apps when it needs them. side work goes to subagents, in parallel.
skills and mcp servers come over from claude code, codex, cursor and claude desktop in one click.
commits, pushes and pull requests wait for your ok. every changed file opens as a diff, and a comment on a line goes straight back to the agent.
once the pull request is up it keeps watching. failed checks, conflicts and review comments come back to it, and it fixes them. your folder updates when you merge.
how it gets better every time you use it, so the hundredth conversation goes better than the first.
after every run it writes down what was built, fixed or decided in that folder. the next conversation starts already knowing it, and can look up more when it needs to.
notes about the same file get merged over time. the ones the code proves wrong are retired, so it doesn't act on stale facts.
when you correct it, it writes a lesson in the same turn. lessons go into every later conversation in that folder, or in all of them.
some lessons are checks. the next time it reaches for the same wrong command or file, the call is stopped before it runs, with the lesson as the reason.
from your replies, your notes on its questions and the things you turn down, it learns how you want it to report, ask, plan and deliver.
each point keeps the words that taught it, and you can edit or delete any of them. say something twice and it moves up the list.
when a command fails and it finds the way past it, the working steps are saved. next time they're already there, so it skips the trial and error.
tell it the order you like for a kind of task and it keeps that too, and plans by it from then on. all of it is listed in settings and saved on your computer.
pair it with your computer to check on agents, answer their questions, approve a push or start new work while you're away from your desk.
it also works on its own, on your claude or chatgpt subscription, with nothing running at home.