Build it for me
Switch it on and it reads the repositories you work in, then suggests small improvements — each with a reason, a plan and a size. Press Go and a chat builds it and opens the pull request.
Hand it the errand you keep putting off, or the bug you never get to. It works on a computer of its own — browsing, signing in, writing code — and tells you when it's done.
Sign in with GitHub, Google, a passkey or a password. Pay only for what you use.
Universal sign-in New
Most agents stop at the login page. Yours signs in the way you do — with your passwords and passkeys, the codes texted and emailed to you, and your Google account — everything a sign-in asks for, on any website, without the AI ever holding the real thing.
Any site you can sign in to
For the boring stuff
Each request gets a real browser on a computer in the cloud. It signs in with your saved logins, reads the code that was texted or emailed to you, and works through the steps — while you do something else. Watch it any time, or take over for a CAPTCHA.
For your side projects
Every chat has its own Linux sandbox with your repositories cloned in. It runs your tests, opens a pull request on GitHub or Forgejo, and you review and merge it without leaving the chat.
1012 before the server had sent it. It now waits for the close frame. 200 runs, 0 failures.fix/reconnect-race → mainFor the big asks
Say "build it with a JIT harness" and the agent writes a small program for your request that runs the work through other chats: some in parallel, some feeding into the next, some going back and forth until the work is good enough. You read the program before it runs, or tell it not to bother.
Also in the box
Three things that start where an ordinary chat stops.
Switch it on and it reads the repositories you work in, then suggests small improvements — each with a reason, a plan and a size. Press Go and a chat builds it and opens the pull request.
Ask for a page and a chat writes one, live at its own address the moment it is done: charts that read your connected APIs, kept private to you, run in a locked-down frame.
The APIs your chats call — yours or the marketplace's — work from any script or app too. Make a client secret, limit it to the APIs it needs, and rotate or revoke it whenever you like.
You decide what it can touch
A new chat starts from a profile you set up. Grant it one site's logins, your mailbox read-only, or a single repository — and take it back mid-task if you change your mind.
Logins, mail, texts, repositories and tools are granted chat by chat, read-only or full, and checked on every call.
Every chat runs in an isolated sandbox, away from your devices. Turn its internet off and it can still reach its model and nothing else.
Every chat's browser is live. Watch it work, and take the controls for a CAPTCHA or a sign-in only you can finish.
Save a set of permissions — "errands", "this repo only" — and start any new chat from it in one tap.
Run chats on your own machine instead of ours, or send its web traffic out from your home connection.
Download everything, or delete your account and all of it, from Settings at any time.
Pricing
You are charged for the minutes a chat's computer is running. When a chat goes quiet, its computer is shut down.
Pay as you go
Top up when you like and spend as you go. Every purchase comes with a receipt.
$10 a month
Usage included every month, for when you use it most days. Credits only once it runs out.
$20 a seat a month
Pro for everybody, with the usage pooled. From 2 seats.
$45 a seat a month
Everything in Team, with more usage per seat. From 10 seats.
Bring your own keys on any plan: use your Claude subscription or your own API keys for the model, so model costs go straight to your provider.
Refer a friend: when they sign up through your link and first pay $10, they get $10 in credits and you get $20.
Make an account and ask it. It can read the codes texted to you, too, so it can sign in where a site sends one — through the Android app, or for an iPhone, a Mac signed in to your Messages.