Guide

Claude computer use: how it works, its limits and alternatives

Where it runs, how to turn it on, what Anthropic says it will not do, and what else you can try.

Summary

Claude computer use is a good default if you already pay for Claude and want no extra app. Anthropic itself calls it the broadest and slowest way Claude can work, with no sandbox between Claude and your apps. The alternatives differ in where the agent runs, how you tell it what to do, and whose model does the work.

Key points

  • Claude Desktop: Pro, Max and Ultra on macOS and Windows. On Team and Enterprise it is a beta that stays off until an owner turns it on.
  • Claude Code command line: macOS only, research preview, not on Team or Enterprise.
  • Claude asks before it uses each app. Browsers are view only, and terminals and code editors are click only.
  • Alternatives: ChatGPT Work and Codex, Manus Studio, Hermes Agent, Fazm, Lapu and Denker.
  • Denker starts from your screen: you point, draw and speak, and your own Claude Code, Codex or OpenCode does the work.

Claude computer use lets Claude see your screen and click and type in the apps on your computer. As of November 20, 2026, Anthropic documents it in three places: the Claude Desktop app on macOS and Windows for Pro, Max and Ultra plans (in chat, Cowork and Claude Code sessions), the Claude Code command line on macOS as a research preview, and a computer use tool on the Claude API for developers. This guide explains how each one works, how to turn it on, the limits Anthropic states, and six alternatives. Denker wrote it, so read it as a vendor's view. It is based on each vendor's own pages, checked November 20, 2026; we did not run a comparative benchmark.

What is Claude computer use?

Computer use is a way for Claude to work in apps that have no connector. Claude takes screenshots to see the screen, then clicks, types and scrolls the way you would. Anthropic's examples include testing a native app, filling in a tool that has no API, and reproducing a layout bug.

It is a last resort by design. Anthropic's help page says Claude picks the most precise tool first: a connector, then the browser, then the screen. The Claude Code docs call computer use "the broadest and slowest" of those routes.

This guide covers Claude's own product. For the general idea behind every agent of this kind, read our explainer on how computer-use agents work.

Where is Claude computer use available today?

There are three surfaces, and their rules differ. Check the one you plan to use.

Claude computer use by surface · Anthropic's own pages, checked November 20, 2026
SurfacePlatformsPlansStatus
Claude Desktop app (chat, Cowork and Claude Code sessions)macOS and WindowsPro, Max and UltraThe Claude Code desktop docs call it a research preview
Claude Desktop app on Team and EnterprisemacOS and WindowsTeam and Enterprise, once an owner turns it onBeta, rolling out gradually; not in Claude Code or tasks started with the Cowork option
Claude Code command linemacOS onlyPro, Max and Ultra; not Team or EnterpriseResearch preview; interactive sessions only
Computer use tool on the Claude APIAn environment you build and runYour Claude API usageFor developers; your app carries out each action

How do you turn on Claude computer use?

In the Claude Desktop app: update the app, then open Settings. Go to Capabilities, or to This computer and then System if your Settings menu lists it. Turn on "Enable computer use". On Windows the toggle works at once. On a Mac you also grant two system permissions: Accessibility, so Claude can click and type, and Screen Recording, so Claude can see the screen. Then start a session and ask for a task in an app. Claude asks before it uses each app.

On Team and Enterprise: an owner of the organization turns it on first. Anthropic says it then works in the new Claude experience only, and that owners can limit which apps Claude may use.

In the Claude Code command line: run /mcp in an interactive session, select the built-in computer-use server and choose Enable. The setting is saved per project. The first time Claude uses the screen, macOS asks for the same two permissions. You need to be signed in through claude.ai; Anthropic says it is not available through third-party providers such as Amazon Bedrock.

On the Claude API: there is no switch. You add the computer use tool to a request, and your own code takes the screenshots and performs the clicks. Anthropic's docs say this needs a sandboxed environment and recommend a dedicated virtual machine or container.

What does Claude ask before it acts?

Claude asks for permission the first time it needs each app. You choose to allow it for the session or deny it. Apps with wide reach, such as terminals, Finder and System Settings, show an extra warning in that prompt.

What Claude may do then depends on the kind of app, and you cannot change these tiers. Browsers and trading apps are view only. Terminals and code editors are click only: Claude can click and scroll, not type. Most other apps get full access. The help page lists music, video streaming and e-book apps as off-limits.

On macOS 15 or later, Claude works in background windows by default and leaves your pointer and keyboard alone. It asks the first time a task needs the full screen in each session. In the command line, pressing Esc stops it at once.

What are the limits of Claude computer use?

These limits come from Anthropic's own pages. None is our test result.

  • It is the slow route. Anthropic says screen interaction is slower and more error-prone than connectors, and that complex tasks may need a second try.
  • There is no sandbox. The help page says: "Computer use has no sandbox between Claude and your applications." It runs on your real desktop, in the apps you approve.
  • Your computer must be awake and the Claude Desktop app open.
  • It does not click or type in your browser. For web work Claude uses the browser built into Claude Desktop or Claude in Chrome.
  • One app can affect another. Anthropic's example: a link clicked in an email may open in Chrome even if you did not approve Chrome.
  • Anthropic advises against using it with sensitive data such as health, financial or legal records. Claude is trained to avoid trading and typing passwords or card numbers.
  • Screenshots can include private information on your screen.
  • In the command line: macOS only, one session at a time, and your other apps are hidden while Claude works.
  • Background windows are documented for macOS 15 or later. Anthropic's pages do not state them for Windows.

On the API, the cost is yours to manage. Anthropic's docs say each screenshot uses roughly 1,000 to 1,800 input tokens, and long tasks collect many of them.

What are the alternatives to Claude computer use?

Six tools document a similar ability on their own pages. The table compares what each vendor says, not measured results. No vendor page tells you how often a click lands in your app, so test the one task you need.

Claude computer use and alternatives · each vendor's own pages, checked November 20, 2026 · not benchmarked
ToolRuns whereHow you tell it what to doModel and accountPlatformsPrice
Claude computer useYour own desktop, in apps you approveChat in Claude Desktop, or the Claude Code command lineClaude, on your Claude planmacOS and Windows (command line: macOS)Part of Pro ($20 a month, or $17 billed yearly), Max ($100) and Ultra ($200)
ChatGPT Work and CodexYour own desktop, through the ChatGPT desktop appChat in the ChatGPT desktop appOpenAI's models, on your ChatGPT accountmacOS and Windows, in supported regionsPart of ChatGPT; OpenAI's computer use page names no plan
Manus Studio (My Computer)Planned in Manus's cloud; commands run on your computerChat in Manus Studio, also from an authorized phoneManus, on your Manus accountApple Silicon Macs, Windows 10 and 11Free download; tasks use Manus credits
Hermes AgentYour own desktopChat with HermesAny tool-capable model you choosemacOS, Windows and LinuxOpen source (MIT); you bring the model
FazmYour own MacChat and push-to-talk voiceClaude Code, Codex or Gemini CLI, on your own accountmacOSOpen source; its site states no price
LapuYour own computer, plus its own browserChatHosted models included in the plan; no API keyWindows 10 or later (64-bit), Apple Silicon Macs on macOS 14 or laterFree to start; paid plans
DenkerYour own desktop and your signed-in browserPoint, draw, speak or type over any appClaude Code, Codex or OpenCode, on your own loginmacOS and WindowsFree: 25 on-screen requests a month. Pro: €19 a month

OpenAI says ChatGPT asks before it uses each app and cannot automate terminal apps or ChatGPT itself. On macOS it can run a scoped task in the background. On Windows it works in the foreground.

Manus says its Computer Use works with apps and your browser in its own workspace, and that you choose whether to allow each app. Its default mode approves safe actions for you and blocks high-risk ones.

Hermes Agent says it acts without moving your cursor, and that clicks and typing need approval unless you turn approvals off. Fazm says it reaches Mac apps through accessibility and Chrome through an extension. Lapu says its Basic mode asks before risky or outside actions.

How is Denker different?

With Claude computer use you describe the task in a chat. With Denker you start from the thing itself. You point at the email, the cell or the button on your screen, draw on it if that helps, and say what you want. Denker sends a screenshot with your marks and your words to the agent.

The agent is the one you already use: Claude Code, Codex or OpenCode, with your existing login. Provider charges are separate from Denker's plans. Denker then clicks, types, picks menus and fills fields in your desktop apps, and reads and acts in tabs of your own signed-in browser through its extension.

Denker asks before it moves your mouse, the first time an agent uses your browser, and before any payment, purchase, transfer, installer or app launch. Sends, deletes and posts are not gated, by design. A card on screen shows each task live, and Stop halts any agent input.

On a Mac, Denker can type into or open things in another app while you keep working. On Windows it works in the foreground.

Edited native recording with fictional data: a Google Sheets report and customer question become an unsent Gmail reply. Review the figures and missing information before using the draft.Full Sheets-to-Gmail walkthrough →

When is plain Claude computer use enough?

Stay with Claude computer use if you pay for Claude Pro, Max or Ultra, do not want another app, and are happy to describe each task in a chat. It also fits developers who want Claude to build an app and then click through it in the same session.

Denker is not the best fit in a few cases. It is built for people who already use Claude Code, Codex or OpenCode, so it is a second tool, and your AI provider's charges stay separate from Denker's plans. It runs on macOS and Windows only; for Linux, look at Hermes Agent. It does not stop before a send, a delete or a post, so read what you ask for. Its direct actions for Mail, Messages, Calendar, Reminders and Notes exist on Mac only, and it does not work in Firefox.

If your work is mostly on the web and you live in ChatGPT, start there. If you want an agent to work without a coding tool or an API key, Lapu documents that.

How do you test one on your own work?

Pick one real task in one app, such as moving three figures from a report into a reply. Give each tool the same task. Then check the app itself, not the chat answer. Count how often you had to step in. This is a suggested test, not a result from this article.

Frequently asked questions

Is Claude computer use safe?

It has safeguards and real risks. Claude asks before it uses each app, some app types are view only or click only, and Claude checks actions for prompt injection. But Anthropic says there is no sandbox between Claude and your applications, calls the capability still early, and advises against using it with sensitive data.

Which plans include Claude computer use?

As of November 20, 2026, Anthropic lists Pro, Max and Ultra. On Team and Enterprise it is a beta that stays off until an owner of the organization turns it on. In the Claude Code command line it is not available on Team or Enterprise.

Does Claude computer use work on Windows?

Yes, in the Claude Desktop app. The Claude Code command line supports it on macOS only; Anthropic tells Windows users to use the Desktop app. Background windows are documented for macOS 15 or later, not for Windows.

Is Claude computer use the same as Cowork?

No. Cowork is one place to work in the Claude Desktop app. Computer use is an ability you turn on in Settings, and Claude can then use it in chat, Cowork and Claude Code sessions.

Can Claude computer use control my browser?

Not by clicking. Anthropic says Claude can see browsers through computer use but cannot click or type in them. For web tasks it uses the browser built into Claude Desktop or your own Chrome through Claude in Chrome.

Why is Claude computer use slow?

It works from screenshots, one step at a time. Anthropic calls it the broadest and slowest route and has Claude try connectors, shell commands and the browser first.

What is the best alternative to Claude computer use?

It depends on how you work. ChatGPT Work and Codex fit ChatGPT users, Hermes Agent fits Linux and your own model, Lapu needs no coding tool, and Denker fits work you start by pointing at your screen with your own Claude Code, Codex or OpenCode.

Were these tools benchmarked?

No. This Denker-authored guide compares each vendor's own pages, checked November 20, 2026.