Claude Computer Use

The URL this listing carries is a documentation page: computer use is a Claude capability you call through the Messages API as a tool, with no plan and no price. The mechanics are a loop you build. Claude receives a prompt and returns tool_use blocks - screenshot, left_click, type; your application executes them in a sandbox you provide, the reference environment in the docs being a Linux desktop under Xvfb; results go back as tool_result blocks, and the cycle repeats until Claude stops asking for a tool or you hit the iteration ceiling.

Cost comes first. You pay for tokens, each screenshot is input at roughly 1,000 to 1,800 of them, and long loops accumulate fast; the maximum-iteration safeguard exists, the docs say, because runaway loops produce unexpected API bills. Older tool versions stay available under beta headers where the current toolset does not apply, so an integration can sit on an interface the vendor has outgrown. The security notes are blunter than most vendor copy: run it in an isolated VM, keep it off sensitive data, restrict internet access, require human confirmation for consequential actions, and assume prompt injection is live, because the docs state Claude can follow instructions found in a page or an image even against your own, and that classifiers do not remove the precautions. This suits a team building its own operator under its own IAM, hosting its own desktop. It is not a subscription, and not a demo for a non-technical colleague.