a feature from Anthropic that lets Claude models operate a computer by viewing the screen and controlling the mouse and keyboard
Anthropic offers computer use in two forms, aimed at two very different users.
Generally available as a toolset of screenshot, click, type, scroll, and similar actions. Your team supplies the rest: a sandboxed display, the apps, the code that turns actions into clicks, and the loop. Billed by tokens.
Anthropic's desktop assistant for macOS and Windows, on all paid plans. Computer use inside Cowork is a beta for Pro and Max plans. Claude asks permission per app and navigates from screenshots while the computer is awake and the Claude app is open.
Both are good at real things: custom agents for engineering teams, and one person's non-repeating work across email, documents, calendar, and screen.
A great model is one part of a recurring procedure. The rest is either your build or a person sitting at the desk.
Anthropic's own guidance is to run it in an isolated VM, keep credentials out of reach, fence the network, and require a person for consequential actions. Building that is your project.
Scheduled cloud sessions cannot reach reports and bank downloads that live in a signed-in desktop browser.
Tying a reserve balance to the bank statement to the cent should not depend on model judgment.
On Team and Enterprise, local Cowork history stays on each user's computer, outside central admin control.
LaunchAI runs on an enrolled agent PC overnight, with deterministic rules and a screenshot of every action. Its model steps can use Claude. See Product →.
| Claude API computer use | Claude Cowork | LaunchAI | |
|---|---|---|---|
| Built for | Engineers building their own agent | One person's work on one desktop | Recurring back-office procedures across a team |
| Who builds | Software engineers | The user | The process owner |
| Where it acts | The environment your team builds | The user's desktop, while awake and open | The user's machine or an enrolled agent PC |
| Runs unattended on a schedule | If your team builds a scheduler | Cloud tasks, away from desktop apps | Schedules, triggers, and a readiness check |
| Business rules | Your code or the prompt | The prompt | Rules and Tables, no model call |
| Human review | A screen your team builds | Per-app permission prompts | A review step anywhere in the workflow |
| Evidence | What your code stores | Conversation history | Before-and-after screenshots of every action |
| Model choice | Claude | Claude | Claude, OpenAI, Azure, Bedrock, or self-hosted, per step |
| Cost unit | Tokens per screenshot and action | Plan seat | License, plus your model account |
LaunchAI's model steps can run on your Anthropic account, directly or through Amazon Bedrock. Deterministic steps, like the reserve tie-out, make no model call. Then point Cowork at the finished packet folder to draft the manager's narrative from numbers that are already reconciled.
Not with computer use. It needs the desktop awake and the Claude app open. Cowork's scheduled tasks run as cloud sessions, which cannot reach apps inside a signed-in desktop browser.
No. It is a capable component. The display, loop, credentials, approvals, scheduling, and evidence store are yours to build.
As of September 2026, Anthropic lists it as a beta for Pro and Max only.
Yes. Model steps send only the screen region being read to the account you configure.
No prompts. No babysitting. No Integrations.
