How to Make Claude Use Your Computer

You can make Claude use your computer by enabling computer use through Claude.ai or by building a workflow with the Anthropic API. In both cases, the session starts with your instruction, Claude sees the screen, takes actions, and stops when the task is done or when you interrupt it.

Before you start, it helps to know what you want Claude to do and what it should not touch. A tool like fenbs lets you record approved task boundaries so the same constraints apply across sessions without retyping them each time.

The two routes for computer use

The first route is Claude.ai. Anthropic has made a computer use beta available to subscribers on supported plans. From Claude.ai you can start a computer use session, give Claude a task, and it will interact with your screen directly in the browser-based interface.

The second route is the Anthropic API. Developers can build their own tools and products that use the computer use tools to give Claude access to a sandboxed or live computer environment. This route requires coding but gives you more control over the setup, including containerized environments that keep your main system separate.

Check Anthropic's computer use documentation↗ (opens in a new tab) for current plan requirements and platform support before you begin.

Write a task that Claude can follow

The most common reason a computer use session goes wrong is a vague instruction. Claude decides what to do by looking at the screen and interpreting your request. The more specific your task, the less it has to guess.

• Name the starting state: "The spreadsheet is already open in Excel on Sheet 2."
• Describe the exact action: "Copy the value in cell C12 and paste it into the Name field in the form on the next window."
• Set a stopping condition: "Stop after you click Submit. Do not navigate away from the confirmation page."
• Flag what needs your approval: "If a dialog asks for a password, stop and let me handle it."

Permissions and what Claude can see

During the session, Claude takes screenshots of whatever is visible on screen. It does not have access to your file system, clipboard, or running processes beyond what is on screen. Anything visible is potentially readable, so close private browser tabs or documents before starting.

On a sandboxed API setup, you can limit what the environment contains. On Claude.ai, Claude uses your live screen, so you manage visibility yourself by closing applications you do not want in scope.

Setting review points

For any task with irreversible steps, add a review point before that step. You can instruct Claude to pause and describe what it is about to do before clicking a confirmation, submitting a form, or making a purchase. Review the description and respond to continue or stop.

Recording these checkpoints in a shared task record means you have the same review requirements next time without rewriting them. In fenbs, Decisions and rules let you write a rule such as "Always pause before any form submission and describe the action" that a connected assistant reads first.

What to do when something goes wrong

In Claude.ai, send a message to interrupt the session. Claude should pause and respond. In an API workflow, use the stop mechanism your implementation provides. If the session has already taken an unwanted action, address the consequence directly rather than restarting: undo the action, restore the previous state, then update your task instructions to prevent it next time.

When computer use is not the right approach

Computer use is slower and less reliable than a proper API or automation tool for the same task. If the application you want to automate has an API, use it. A Python script calling a REST API is faster, more predictable, and easier to debug than an agent clicking through a GUI.

Computer use is also not suited for tasks that require reading or acting on genuinely sensitive information, such as personal health data, financial credentials, or private communications. The session takes screenshots of whatever is on screen; anything you would not want photographed should stay out of the session.

For code-related work on your machine, Claude Code is a better fit than computer use. It operates in the terminal with direct file system access and does not depend on reading a GUI. If your task is about reading and editing files, running tests, or working with code, start there rather than with computer use.

How to do this with fenbs

1. Create a free account and add a task for the computer use job.
2. Write the task note with the starting state, the goal, and where the session should stop.
3. Add Decisions and rules for the constraints: off-limits areas, approval checkpoints, the ending condition.
4. Connect Claude over MCP from Integrations. At the start of each session, ask Claude to read your project context and rules first.
5. After the session, record what was done and any edge case that came up, so the next run starts with that information.

For more on what Claude can and cannot do with computer access, see can Claude control your computer. If you are using Claude for code work alongside computer actions, what is Claude Code covers that route.