Creativity · Agent Protocol
Anthropic Computer Use Agent
Computer Use, shipped by Anthropic in October 2024 for Claude 3.5 Sonnet and matured across Sonnet 4.x, exposes a small set of API tools — screenshot, mouse_move, left_click, type, key — that together let the model drive any GUI. It was the first production-grade 'computer-using agent' from a frontier lab, designed to let developers automate previously-unscriptable desktop and web workflows.
Protocol facts
- Sponsor
- Anthropic
- Status
- stable
- Spec
- https://docs.anthropic.com/en/docs/build-with-claude/computer-use
- Interop with
- OSWorld, browser-use, Claude Agent SDK
Frequently asked questions
Does Computer Use run locally?
The model runs on Anthropic's servers. The 'computer' it controls is whichever desktop your client exposes — typically a sandboxed VM you operate. Anthropic strongly recommends isolation from sensitive systems.
What about OSWorld scores?
Claude Sonnet 4.x with Computer Use is among the top scorers on OSWorld as of early 2026, but still well below human performance. Reliability, not capability, is the main frontier.
Is there prompt-injection risk?
Yes — screenshots can contain hostile text that the model might follow. Anthropic publishes guidance on detection and sandboxing; production deployments require careful permissioning and supervision.
Sources
- Anthropic Computer Use docs — accessed 2026-04-20
- Introducing computer use — Anthropic news — accessed 2026-04-20