A computer use agent is an AI agent that operates a computer or browser the way a person does. It looks at the screen, usually through screenshots, or reads the web page directly, decides on the next step, clicks and types, and keeps going until the task is done. That helps wherever there is no interface to connect to: supplier portals, older programs, forms. Anthropic offers this as a tool in its API, in Claude in Chrome and in Claude Cowork; OpenAI as Computer Use in the ChatGPT desktop app and with ChatGPT Dots, which has its own computer in the cloud. Open-source options include Nod by n8n and OpenClaw. The catch: an agent like this reads everything on the screen, including injected instructions, and it acts with your access. Where an interface exists, a fixed workflow is usually faster, cheaper and more reliable. All vendor details as of November 15, 2026.
My workforce has a rule: no credentials handled by AI. When Sebastian, my SEO employee, was supposed to open the AI report in Google Search Console with Claude in Chrome on September 15, 2026, the extension took screenshots and read the page content. At the login, he stopped and reported back instead of signing in. I installed the Nod server from source on my Mac on November 15, 2026, version 0.7.6. Everything else is based on docs and help pages from Anthropic, OpenAI and n8n, retrieved on November 15, 2026. Everything I have written about AI agents is collected on my AI agents page.
What is a computer use agent?
A chatbot answers. A workflow runs fixed steps that you built in advance. A computer use agent gets a goal and a computer and finds its own way: it opens a page, sees a form, fills it in, checks the result and moves on. Anthropic describes its tool as screenshots plus mouse and keyboard control for autonomous work on a desktop. OpenAI writes that with Computer Use, ChatGPT can see and operate graphical user interfaces on macOS or Windows.
When the work stays in the browser, people often call it a browser agent. The difference is reach: a browser agent operates websites, a computer use agent also desktop programs, files and system settings. What an AI agent is in general is covered in What Is an AI Agent? Explained Simply, and how agentic AI differs from generative AI in Agentic AI Explained: What It Is and Isn't.
How a computer use agent works
Anthropic explains the process in four steps, and at its core it looks the same with the other vendors:
- Task and tool. The model gets your task and permission to operate the computer.
- The model picks an action. For example: take a screenshot, click a spot, type text.
- A program carries it out. It really clicks, takes the new screenshot and sends the result back to the model.
- Repeat. The model looks at the result and decides whether anything is left to do. If not, it replies.
Anthropic calls steps 3 and 4, repeated without your input, the agent loop. The key point: the model does not sit inside the computer. It only sees what the program shows it and can only do what the program carries out.
Screen or page. There are two ways of seeing. One works with screenshots and coordinates, like Claude computer use in the API. The other reads the web page itself; Anthropic has a separate browser tool for this that doesn't need a full desktop. Claude in Chrome does both: the extension reads the page text and takes screenshots.
Where the computer is. This decides the risk and where data goes:
- a virtual machine or container that you provide. Anthropic's example is a Linux desktop with Firefox and LibreOffice in a Docker container.
- your own computer, as with Computer Use in the ChatGPT desktop app or in Claude Cowork. On a Mac, the ChatGPT app needs Screen Recording and Accessibility permissions for this.
- a computer in the vendor's cloud, as with ChatGPT Dots or Cowork sessions on Anthropic's servers.
- a container on your own computer or server, as with Nod.
Cost. Every screenshot goes to the model as an image and is billed as image input with API pricing. The Anthropic computer use tool alone adds about 4,500 input tokens per request, according to the docs. Many small steps add up.
Computer use agents at a glance
Anthropic: computer use, Claude in Chrome and Cowork
Claude computer use is a tool in the API. One entry in the request gives Claude 17 member tools, for screenshots, clicks, typing and zooming among others; your application carries out each call in an environment you control. Anthropic recommends a dedicated virtual machine or container with minimal privileges, no access to login details, an allowlist of domains, and confirmation by a human before anything with real-world consequences, such as accepting cookies, paying or agreeing to terms. According to Anthropic, screenshots and files are captured and stored in your environment; every screenshot still goes to the model with the request. As limits, Anthropic itself names speed, mistakes with coordinates and lower reliability with niche programs.
Claude in Chrome is an extension that lets Claude read, click and navigate websites alongside you, from the side panel, from Cowork or from Claude Code. It only runs in Chrome, not in other Chromium browsers and not on mobile. Anthropic itself writes that the extension is enhanced with safety classifiers but "still risky". How to set it up is in Claude in Chrome: How to Use the Extension.
Claude Cowork runs tasks on Anthropic's servers in an isolated environment. For websites, Claude uses a browser in the desktop app or your Chrome. With computer use, in beta on Pro, Max and Ultra, Claude operates programs on your computer by clicking, typing and navigating, but only while Claude Desktop is open. What else Cowork does is in Claude Cowork: What It Does, Who It's For.
OpenAI: Computer Use, ChatGPT Work and Dots
Computer Use in ChatGPT is available in supported regions in the desktop app for macOS and Windows, with ChatGPT Work and Codex. ChatGPT asks before it uses an app; with "Always allow" it can use that app without asking in the future. On Windows it takes over the active desktop, meaning the pointer and the foreground, while the task runs. OpenAI's own advice: if an app has its own plugin or MCP server, that structured integration is the better choice for data access and repeatable operations.
ChatGPT agent and ChatGPT Work. The ChatGPT agent page now redirects to ChatGPT Work. According to OpenAI, Work uses your apps, files, tools and a built-in browser in the desktop app. How Work and Cowork differ is in ChatGPT Work vs Claude Cowork: Delegating.
ChatGPT Dots is an always-on agent with its own computer and browser in the cloud. It keeps working when your computer is off. If you want to step in, you take over the computer with "Take over". Before certain actions, such as sending an email, "Auto-review" checks the recipient and content against your instructions and rules; whether that covers every browser and desktop action is still under review, according to OpenAI's own docs. Dots is available on Pro and Business Premium, and on Pro not in the European Economic Area, the UK or Switzerland.
Nod: an open-source computer use agent
Nod by n8n is an open-source agent under Apache 2.0 with its own computer: an isolated Linux workspace for commands and its own Chromium browser in a container, on your computer or server. According to the docs, the browser reaches only public addresses, never your local network. The Nod server checks every browser step that can change something against your request; the check sees the action, the website and the button's label, never the page text. The Vault enters logins into the matching field itself, and the model only sees account names and status. You can use any model you like, though Claude only with an API key. What Nod can do overall is in Nod by n8n: What the AI Agent Can Do, and the comparison with Cowork, Dots and Grok Bot in Claude Cowork Alternative: Nod Compared.
OpenClaw
OpenClaw is another open-source agent under the MIT license that you control through messaging apps and that operates the shell, files and a browser. Tools run directly on the machine unless you set up a sandbox. How to secure it is in OpenClaw Alternative for Businesses. The direct comparison with Nod is in Nod vs OpenClaw vs Hermes Agent Compared.
Risks: what can go wrong
Injected instructions. A computer use agent reads everything on the screen. Anthropic writes that in some circumstances Claude follows instructions from web pages or images, even when they conflict with yours. That is why Anthropic runs classifiers over the screenshots, OpenAI trains the models behind Dots not to follow instructions hidden in emails, files and web pages, and Nod doesn't let pages or emails grant permission. No vendor rules the attack out. How it works is covered in Prompt Injection: Protecting AI Agents.
Approvals. The more independently the agent works, the more it matters where it stops. Anthropic names accepting cookies, payments and agreeing to terms as examples for human confirmation. Nod has you approve every payment with your card's security code.
Credentials. Anthropic advises against giving the model login details. Nod keeps passwords away from the model through the Vault, but writes itself that a secret you type into the chat is known to the model, which can enter it into a page. The safest approach is to sign in yourself. On September 15, Sebastian stopped right there and called me instead of signing in.
Mistakes. An agent can misclick, hit the wrong coordinates or take an unexpected shortcut. On your own computer, that can affect things outside your project folder, which is why OpenAI advises clearly scoped tasks.
Data flow. Every screenshot goes to the model provider. Whatever is on the screen, including customer data in an open window, ends up there. What exactly Nod sends to the model, n8n and other services is in Nod Privacy: What the AI Agent by n8n Shares.
When a computer use agent pays off and when an n8n workflow is better
The rule of thumb is already in the vendors' docs: OpenAI recommends the plugin or MCP server over Computer Use for apps that have one, and Anthropic recommends the leaner browser tool for tasks that stay on web pages. How to build a fixed process is in n8n Workflow Tutorial: Build, Test, Export; when to use a workflow and when an agent, in n8n, Make, or Claude: What For?. How Claude reaches services through interfaces instead of the screen is explained in Claude Connectors: What They Are and Do.
Frequently asked questions
What is a computer use agent?
An AI agent that operates a computer or browser itself: it sees the screen through screenshots or reads the page, clicks, types and checks the result until the task is done. A program carries out each action the model proposes.
Is there an open-source computer use agent?
Yes, for example Nod by n8n under Apache 2.0, with its own browser and Linux workspace in containers, and OpenClaw under the MIT license. For Claude computer use, Anthropic provides a reference environment on GitHub that you run yourself.
What is the difference between a browser agent and a computer use agent?
A browser agent only operates websites, often by reading the page directly. A computer use agent operates the whole computer through the screen, including desktop programs and files. Many offerings can do both.
What does Claude computer use cost?
It is billed per token through the API. According to Anthropic, the tool itself adds about 4,500 input tokens per request, plus each screenshot as image input. In Claude Cowork and Claude in Chrome, usage is included in the paid Claude plans.
Is a computer use agent safe?
As safe as the environment and approvals you give it. Vendors filter injected instructions, but none can rule them out. Let the agent work in its own environment, don't give it passwords, and keep payments and anything that leaves the house in your hands.
Where is ChatGPT agent today?
The ChatGPT agent page redirects to ChatGPT Work. Computer Use is available there in the desktop app for macOS and Windows, and an always-on agent with its own cloud computer comes with ChatGPT Dots.
How to continue
Pick a task that is stuck today on a portal without an interface, and write down where the agent has to stop: before the login, before submitting, before paying. Test it first in a separate environment with test data. If there is an interface for the same task, build a workflow instead. How I build AI employees with clear permissions and approvals is something I show in my community Claude Practitioners.
