Fleet is a Mac app. Give it one goal in plain English, and an AI captain puts a crew of AI engineers on it. Nothing comes back done until a check proves it.
36 of 50 free beta seats left.
Fleet is early. Every worker is the real Claude Code, running on the Claude plan you already pay for.
Two of them are yours: one sentence at the start, one yes or no in the middle.
Plain English, with the outcome in it. You bring the intent. The Captain, the AI you talk to, owns the craft and hands the job to a worker.
The Goal and Proven screens come from a stand-in run.
I understand this pushes code beyond this machine.Worker 1 stopped here and waited for one click. Yours, even from your phone.
Worker 1 · You can always deny.Fleet refuses a phone tap on the biggest calls: deleting things, changing your database, putting code live. Those wait for you at your desk.
Proven.A worker does not get to say when it is finished. Fleet ran the tests itself.
Worker 1 · the project’s own tests, passing in worker 1’s own copy.
Where it stalls
AI made starting cheap, so there is more of it. The finishing stayed where it always was, with you: the testing, the fixing, the reading of every change, the part where it actually goes live. Fleet takes the finishing.
It takes your goal, plans the work, designs the checks that will prove it, starts the workers, and judges what comes back. Describe the job the way you would to a person. The notes a new engineer reads first are not yours to write either: if your project has none, Fleet writes them before work starts and sends a worker to fix them when they fall short.
Every job builds in a copy of its own, so one worker cannot overwrite another. Every worker is on screen while it works: send it a new instruction and it picks that up without starting over. You do not have to sit and watch them: something checks on every worker for you, on a timer that uses none of your Claude plan. If one gets stuck, Fleet saves its work before touching it.
A worker cannot mark its own job done. The check has to run and pass first, and a check that is missing, broken, or ran no tests counts as a fail, the same as one that ran and failed. A fail goes back to the worker with the reason, up to three times, and then the job comes to you. When your project lives on GitHub, what passes arrives there as a change offered for your review, kept apart from the project your team works from. Fleet cannot accept it for you, cannot write over that project, and cannot erase work already saved.
A screenshot only proves a design exists. So before design work counts, an AI from a different company reviews the page against a written list of the ways AI-made pages look cheap. A poor page goes back to the worker that built it. If that reviewer is not set up, Fleet tells you the design was not independently reviewed.
Your keys and passwords stay in your Mac’s Keychain, the same place the system keeps the rest of them. The AI never sees anything that unlocks anything, and that holds for every worker in the crew.
every card
Every card carries its own finish line, the check that decides whether its job is done. A worker that goes quiet stays marked working until that check passes.
top right · needs you
This worker wants to send code to your live service. Fleet always stops here, and there is no way to turn that off.
bottom left · the red card
This worker ran out of tries and stopped. Its reason is printed on its own card, at the same size as a pass.
On your own, the work moves at the speed you can sit with it. One job at a time, and you read every change before you trust it. A crew moves several jobs at once, and not one of them comes back until a check has proved it. More gets finished in a week, and each finished thing arrives with the check that proves it.
Fleet runs on your own Claude plan, so a whole crew working all day costs the same flat price. A metered agent is a taxi meter. Your plan is a monthly pass. A crew draws on that plan faster than one worker does, and when you hit its limit they wait and pick back up when it resets.
This app is built by its own crew, and on August 18 that crew shipped a fix for a check that could not fail. The check reported success whatever the tests said, and it had marked two workers done while one of them hid thirteen real test failures. A check that reports success without running the tests now counts as a failure, because Fleet reads the real test results itself. You’ve got a list of things you want built. Fleet is the crew that finishes them, and proves each one before it says done.
JOBPROVEN
Set the dial to Supervised, the position that checks with you, and hand Fleet a job before bed. In line, running, sent back, proven: those happen while you sleep. The judgment call is waiting for you in the morning, and the work Fleet is sure is safe keeps going around it.