Front page / Concepts / Computer use
Computer use
A computer use agent sees the screen and clicks, types and scrolls like a person, so it can work in any website or app that has no connector: filling forms, pulling numbers from a dashboard, checking a competitor's checkout flow. It is slower and less reliable than a proper tool connection, and more general.
Where it stands
Among the 14 concepts. Bars: GitHub stars gained this week.
- MCP and tool use+3.9k
- Agentic loops+1.3k
- Knowledge graphs (GraphRAG)+1.2k
- Retrieval (RAG)+1.2k
- Prompt engineering+1.1k
- Vibe coding+980
- Workflows, not agents+565
- Spec-driven work+16k
- Memory+7.3k
- Agent skills+5.7k
- Multi-agent orchestration+5.6k
- Computer use+4.7k
- Context engineering+3.3k
- Evals and verification+929
How to use it today
In ChatGPT or Claude, as a person doing the work, not as a developer.
- 1Use the browser side of your assistant (ChatGPT Work, Claude in Chrome) for tasks that live in a website without a connector: a form, a portal, a dashboard export.
- 2Watch the first few runs; approve each action that spends money, sends a message or deletes something.
- 3Give exact instructions about what to copy out and in what shape; the agent is good at navigating and poor at deciding what matters.
- 4Prefer a connector (MCP) when one exists for the app; use computer use for the long tail.
- 5Log in yourself first; do not hand the agent passwords.
What changes day to day
Repetitive web work (checking ten portals, filling the same form weekly) leaves your hands. You stop doing it and start checking that it was done.
Where it goes wrong
Pages change and the agent clicks the wrong thing with the same confidence.
Prompt injection hidden in a web page can redirect an agent that has your session.
Listings that use it well
Chosen by hand after reading them, each with the reason. Install from the listing's own page.
| Listing | Kind | Why it is here | Stars |
|---|---|---|---|
| Chrome DevTools MCP | MCP server | Chrome DevTools over MCP for driving and inspecting a live browser. | 55k |
| ai-agent-book | Agent skills | computer-use covers the perceive, think, act loop, three visual grounding routes and coordinate scaling. | 57k |
| agent-browser | Agent skills | Vercel's browser automation CLI for agents, with Electron desktop app control. | 48k |
| mcp-server-browser | MCP server | Browser use access from ByteDance's UI-TARS desktop agent. | 40k |
| playwright-mcp | MCP server | Microsoft's Playwright tools over MCP; the accessibility tree becomes the agent's view of the page. | 39k |
| browser-use | MCP server | Controls a real Chrome to fill forms, extract data and book flights. | 120k |
Standing
Our judgment, with the reason and the sources it rests on.
Rising since 2025-07. Anthropic's computer use beta (October 2024) was for developers only. It reached the chat apps with ChatGPT agent in July 2025 and Claude in Chrome in August 2025. In 2026 OpenAI folded it into ChatGPT Work and Claude in Chrome became generally available on paid plans, with automatic safety checks on each action. Adopted faster each month; the vendors are building for it and new tools keep appearing.
- Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku2024-10-22
- Introducing ChatGPT agent2025-07-17
- Claude in Chrome is generally available2026-08-26
Related: MCP and tool use, Agentic loops, Workflows, not agents