Disclosure: I'm a developer who works with React, Next.js and AI agents, and I offer related services, so I have a business interest in this topic. AI tools helped with research and drafting; I reviewed the content before publishing. Figures are estimates as of November 2026 and may change; check official pricing pages.
Quick Summary
Claude computer use lets Claude see a screen and use a mouse and keyboard. To use it safely at work, give it the least access possible: a sandboxed machine, no stored passwords, an allowlist of websites, and a human approval step for risky actions. Treat the AI agent like a new intern, not an admin.
Key Highlights
- Computer use is a client-side tool: your app runs every click in an environment you control.
- Anthropic's docs list four core precautions: isolated VM, no sensitive data, domain allowlist, human confirmation.
- Webpages and images can carry hidden instructions (prompt injection). Plan for it.
- On the desktop app, Claude asks permission per app, and trading and crypto apps are blocked by default.
- Who should read this: IT leads, founders and developers rolling out agents.
What Is Claude Computer Use?
Claude computer use is a tool that gives Claude screenshots plus mouse and keyboard control. Claude looks at the screen, decides the next click, and repeats until the task is done.
Per Anthropic's docs, the current toolset bundles 17 actions such as screenshot, left_click and type. Your own code runs each action. That matters for security, because you decide where Claude runs.
Two Ways to Use It: Desktop App vs API
The right choice depends on who is building. The desktop app suits individual workers. The API suits teams building controlled agents.
| Claude desktop app (Cowork / Claude Code) | Claude computer use API | |
|---|---|---|
| Runs on | Your own computer | A VM or container you build |
| Permissions | Per-app prompts | Your code and your sandbox |
| Best for | Personal desktop tasks | Business workflows, testing |
| Windows support | Yes (macOS and Windows) | Any OS you can run in a VM |
| Plan status | Beta; see plan note below | Pay per token |
Plan note: Anthropic's help pages show computer use in beta for Pro, Max and Ultra plans. One older-looking help snippet says Team and Enterprise plans lack it. Confirm your plan on the official help page before you promise anything to your team.
How to Set AI Agent Permissions Safely
Start with zero access and add only what the task needs. Here is a simple permission model.
- Isolate the machine. Use a dedicated VM or container with minimal privileges, as Anthropic advises.
- Keep secrets out. Do not give the agent login data for sensitive accounts.
- Limit the internet. Allow only the domains the task needs.
- Add human approval. Ask a person to confirm payments, cookie consent, terms of service and anything with real-world impact.
- Log and cap. Keep action logs and set a maximum step count to stop runaway runs.
Use APIs and connectors first
Screen clicking is the slowest, least precise route. If a task has an API or an MCP connector, use that instead. The Cowork help page says Claude tries the most precise tool first. Computer use is the fallback for apps with no other door.
Claude Computer Use on Windows
Claude computer use works on Windows through the Claude Desktop app, in both Cowork and Claude Code. Windows runs Cowork's code inside a Hyper-V virtual machine, and the app needs the Virtual Machine Platform feature turned on.
IT teams can deploy it with MSIX packages through Intune, SCCM, Group Policy or PowerShell. One warning: Anthropic's help page says computer use has no sandbox between Claude and your apps. Close sensitive files first.
Admin Controls Worth Knowing
Enterprise admins already have useful switches, per Anthropic's Cowork documentation:
- An organization-wide toggle to turn Cowork on or off.
- Groups and custom roles on Enterprise plans, so only chosen teams get access.
- "Automatically approve" mode, which is off by default on Enterprise.
- "Always allow" for write-capable connector tools, also off by default.
If you are planning wider agent rollouts, see my post on why AI agents are replacing SaaS seats in 2026 and browse more in AI agents.
Which Tasks Fit Computer Use (and Which Don't)
Computer use works best when speed is not critical. Anthropic's docs point to background information gathering and automated software testing in trusted environments as good fits.
Avoid it for jobs that need perfect precision, such as payroll runs or legal filings. Also avoid anything touching health, finance or other personal records. Anthropic's help pages advise against using it on apps with that kind of data.
A Simple Permission Tier Model
Not every task needs the same trust. This three-tier model is my own suggested framework, not an Anthropic feature. It helps teams decide how much freedom to give.
| Tier | What the agent may do | Controls to add |
|---|---|---|
| 1. Look only | View screens, read data, report back | Sandbox VM, domain allowlist |
| 2. Work in a sandbox | Fill forms, edit test files, run tests | Action logs, step limit |
| 3. Act on real systems | Submit, send or change live data | Human approval before every action |
The API helps here. The toolset lets you switch off individual actions using its configs setting, so you can withhold any action your workflow does not need.
A Safe 30-Day Rollout Plan
Go slowly. Small pilots catch problems while the damage is still small.
- Week 1: Pick one low-risk task. Build the sandbox and allowlist.
- Week 2: Run it at Tier 1 with a person watching every step.
- Week 3: Review the logs. Fix confusing prompts. Add instructions for common errors.
- Week 4: Move to Tier 2 only if results are steady. Keep approvals for anything risky.
Hidden Costs, Mistakes and Red Flags
Most failures come from over-trust, not from the model itself.
- Prompt injection. Text on a page can trick Claude into following the wrong instruction. Anthropic added classifiers that flag likely injections, but its docs say precautions still matter.
- Passwords in prompts. Giving credentials raises the damage if an injection works.
- No human in the loop. Anthropic notes the extra injection protection fits poorly when nobody is watching.
- Wrong clicks. Claude can misjudge coordinates or pick the wrong tool, especially in niche apps.
- Token bloat. The toolset adds roughly 4,500 input tokens per request, and each screenshot costs about 1,000 to 1,800 more.
Pricing: What Does Computer Use Cost?
There is no separate computer use fee. Anthropic's docs say it follows standard tool-use pricing, so you pay for tokens.
| Cost item | Official figure |
|---|---|
| Toolset overhead | About 4,500 input tokens per request |
| Each screenshot | Roughly 1,000 to 1,800 input tokens |
Long loops add up fast. For budgeting, read my AI agent development cost guide.
FAQ
What is Claude computer use?
It is an Anthropic tool that lets Claude view a screen and control a mouse and keyboard to finish tasks in apps and browsers.
How to use Claude computer use?
Individuals can turn it on in the desktop app settings. Developers add the computer toolset to a Messages API request and run the returned actions in their own sandbox.
Does Claude computer use work on Windows?
Yes. Anthropic's help pages say it is available in Cowork and Claude Code in the Claude Desktop app for macOS and Windows.
Is Claude computer use safe for business?
It can be, with limits. Use a sandbox, an allowlist, no sensitive logins, and human approval for risky steps. Anthropic itself warns against sensitive data without oversight.
Where is my data stored?
Anthropic's API docs say screenshots, inputs and files stay in your environment.
Is it better to use MCP or computer use?
Use MCP or direct APIs when they exist. They are faster and more precise. Use computer use for apps that offer no other access.
Conclusion
Safe enterprise use of Claude computer use comes down to least privilege: sandbox it, limit its websites, hide secrets, and require approval for risky actions. Start with one low-risk workflow, review the logs, then widen access slowly. If you want help designing agent permissions for your team, see my AI & workflow automation service or get in touch.
Written by Amrendra Kumar, Frontend Developer & Technical Content Writer, working hands-on with React, Next.js and AI-agent projects.




