Clerk is an AI that actually uses a computer. Describe the job in a browser tab and it works on your own Mac — moving the mouse, typing, opening pages, reading files, and handing coding work to Cursor and Claude Code.
Pay per use · No subscription · Stops when your balance runs out, never bills you by surprise
Why Clerk
Clerk is not the only agent that can drive a computer any more. These three are where it genuinely differs — two are capabilities, the third is the bill.
Not a browser extension, not a bot that can only call APIs. It sees the screen, moves the pointer, types, and opens whatever you have installed. Your files do not have to reach somebody's cloud first.
Give it an OpenAI-compatible endpoint and the thinking goes to a server you run — vLLM or Ollama on this box, or your company's own gateway. Conversations stop passing through us. A product tied to one model vendor cannot offer this without undercutting itself.
$10 on sign-up, then only what you use; the balance does not expire. Each finished task tells you what it cost — usually cents. No monthly fee, and never "you have used this month's tasks".
How it works
No workflows to configure, no scripts to write. Explain it the way you would to a colleague.
Download and open it. The first launch asks for Accessibility and Screen Recording permission — it needs to see the screen and move the mouse.
The interface is a chat box running on 127.0.0.1, on your own machine. You can talk to it instead of typing, or open phone mode and just speak.
What it saw, where it clicked, what it plans next — all streamed live. Interrupt whenever you want; it picks up your new instruction on the spot.
Conversations follow your account. Start on your laptop, sign in on your desktop, keep going.
What it does
Not a browser extension, not a bot that can only call APIs. It has hands.
It screenshots the screen, then genuinely moves the pointer, clicks, types and scrolls. Whatever is installed on your Mac, it can drive.
On a web page it numbers the clickable elements first, then says "click 7" instead of guessing coordinates. That removes a whole class of near-miss errors.
It doesn't grind through coding itself. It hands the sub-task to a dedicated coding agent and collects the result.
Tap the mic and talk — words appear as you speak. Phone mode keeps the line open so you can work hands-free and cut in mid-sentence.
Reads and writes files, runs commands, looks things up online. What you can do on your Mac, it can mostly do too.
Deleting, overwriting, reading keys, changing system settings, discarding uncommitted work, sending local files out — those stop and ask first, naming the consequence and the path rather than handing you an error afterwards. Ordinary commands do not interrupt you.
Long jobs get a to-do list it works through. When a step goes wrong it notices and retries without being told.
It speaks MCP, so Notion, Linear, Slack, Sentry and Zapier connect in one click and hosted services open their own sign-in page. You can also write your own procedures and roles for it to follow.
Pricing
Top up a balance, spend it as you go. No monthly fee, no auto-renewal, and when the balance is gone the service stops — you can never end up owing money.
$10 credit
$0 / month
| Model | Good for | Input / M tokens | Output / M tokens |
|---|---|---|---|
| Qwen3.8-27B · default | Everyday work, best value | $0.63 | $3.83 |
| GLM 5.3 Flash | Cheapest that can still see a screen | $0.11 | $0.38 |
| Claude Sonnet 5 | Harder work | $3 | $15 |
| Claude Fable 5 | The hardest GUI tasks | $15 | $75 |
Thirteen models to choose from, across Anthropic, OpenAI, Qwen and Zhipu, switchable in the window. Three honest notes. GUI work costs more than text work — every step sends a screenshot to the model, so a multi-step desktop task is usually tens of cents on the default model, while a text-only task is often a few cents. The cheapest model per token is not always the cheapest per task: on our own benchmark Qwen3.8-27B finished every run at a fifth of the price, but needed about a quarter more steps to get there. And none of this column applies if you bring your own model — then you pay for your own compute rather than for ours (voice will need a key for a speech service of your own).
Download
On a Mac, pick the chip — Apple menu → About This Mac.
Two honest notes about the Windows build. It is not code-signed yet, so the first run brings up SmartScreen's "Windows protected your PC" — More info, then Run anyway. If you would rather verify the file than trust that dialog, the SHA-256 of every published build is at /version; check it with Get-FileHash. And in-app updating is macOS-only for now — on Windows the app tells you a new version exists and you unzip it over the old folder.
There is no iOS equivalent, and that is not for want of building one: iOS has no third-party accessibility service, and no public API lets an app read another app's screen or inject touches. What an iPhone can do is drive your computer — open this page in its browser.
Huawei laptops that came with Windows (MateBook and friends) run the Windows build above. HarmonyOS PCs are not supported — they run a native HarmonyOS kernel, do not run Windows or Linux applications and do not allow sideloading, so they need a separate native app we have not built.