There are two ways to let an AI agent use software that was built for people. A computer use agent looks at screenshots of a whole desktop and moves the mouse and keyboard. A browser use agent drives a real browser directly: it reads the page, acts on its elements and talks to the browser over the Chrome DevTools Protocol (CDP).
Both can finish a task like "download last month's invoice from the billing portal". They get there in different ways, and that decides how fast, how reliable and how expensive the agent is, and what it can reach at all.
TL;DR
Use browser use when the work happens on websites, which covers most of what agents are asked to do. Use computer use when a step leaves the browser: a desktop app, an operating system dialog, or software with no web version. Many workflows need both, with the browser doing the web part.
| Compare | Browser use | Computer use |
|---|---|---|
| Controls | A browser: tabs, pages, cookies, network | A whole desktop: any app, any window |
| Model sees | Page structure and text, plus screenshots when they help | Screenshots |
| Model acts with | Clicks and typing on elements, JavaScript, CDP commands | Screen coordinates and keystrokes |
| Runs on | A local or cloud browser, often headless | A virtual machine with a desktop, or your own screen |
| Good at | Precise, scriptable, parallel web work | Anything a person can see |
| Weak at | Anything outside the browser | Small targets, long pages, anything off screen |
What is computer use?
"Computer use" is the name Anthropic gave Claude's ability to operate a computer from screenshots, announced in October 2024. OpenAI and Google have since shipped models that work the same way.
The loop is simple. The harness takes a screenshot and sends it to the model. The model answers with an action such as "click at this position" or "type this text". The harness performs it, takes a new screenshot, and the loop repeats until the task is done.
The appeal is that it works on anything with a screen. The model needs no API and no knowledge of how the app was built.
The cost is that the model has to find everything by looking. It only sees what fits on the screen, so long tables, dropdowns and pages have to be scrolled into view piece by piece. Every step sends a new image to the model, which adds time and tokens. Small controls make precise clicks harder. And the agent needs a full desktop to run in: a virtual machine, or your own computer, which it takes over while it works.
What is browser use?
A browser use agent drives a browser instead of a screen. It reads the page as data: the document structure, the accessibility tree, the text, links and form fields. It acts on elements rather than coordinates: press this button, fill this input, run this script, read that network response. It connects over CDP, the same protocol Playwright and Puppeteer use.
The name is ours. We launched Browser Use as an open-source library in November 2024 so any model could control a browser. Today Browser Use is:
- The open-source library (browser-use, MIT), for running your own agent on your own Chrome or on a cloud browser.
- Browser Use Cloud, stealth browsers you connect to from your own code or agent, with CAPTCHA solving and residential proxies on by default.
- Browser Use Agents, a hosted agent: send a task, get the result.
- Browser Harness, which lets coding agents like Claude Code and Codex drive your own Chrome.
Why browser use fits web tasks better
The page already describes itself. A button has a label and a role. A form field has a name. The agent can target them directly instead of guessing where they are on screen, and it can reach elements that are scrolled out of view.
The model sees the whole page at once. Reading a page as text lets the model take in a full table or form, not just the part that fits in a screenshot.
Deterministic steps can be code. When part of a task is predictable, the agent can run a script instead of reasoning through each click. Pulling a table out of a page is one JavaScript call, not a dozen scroll-and-read steps. We wrote about how far this goes in The Bitter Lesson of Browser Agents.
Browsers scale. A headless browser needs no desktop, so you can run many at once, each isolated from the others. On Browser Use Cloud a browser costs $0.02 per browser-hour with no subscription (pricing).
The hard part of the real web is the browser, not the model. Bot protection, CAPTCHAs, proxies, logins and session state decide whether an agent gets in at all. Those are browser infrastructure problems. We publish a stealth benchmark of cloud browsers on bot-protected sites, with the code and raw results.
When computer use is the right call
Computer use earns its cost when the task cannot be done in a browser:
- Desktop apps with no web version, such as older Windows business software or native design tools.
- Operating system steps: native file pickers, system settings, installers, moving files between apps.
- Pages where the structure says nothing, such as a remote desktop session or an app drawn entirely on a canvas. Here the pixels are the only interface.
The last case is narrower than it sounds. A browser use agent can take screenshots too. Vision is one of its tools, used when the page structure runs out, not the only way it can see.
Use both
Real workflows often cross the line. An agent might pull a report from a web dashboard, then open it in a desktop app. The rule of thumb: do the web part through the browser and keep pixels for what only pixels can reach.
The same goes for agents on your own machine. Computer use moves the mouse on your screen, so you wait while it works. A browser agent can work in a Chrome tab in the background while you keep using your computer. That is how Claude computer use in your browser works with Browser Harness.
How to choose
- Is every step on a website? Browser use.
- Do you need many sessions in parallel, logged in or on bot-protected sites? Browser use on cloud browsers.
- Does a step need a desktop app or an OS dialog? Computer use for that step.
- Do you already have Playwright or Puppeteer code and only need a remote browser? That is not an agent question. Point it at a cloud browser: see the Playwright guide and how cloud browser providers compare.
If you are choosing between cloud browser providers, we keep sourced comparison pages for Browserbase, Browserless, Steel, Kernel, Hyperbrowser and Anchor, each with prices from the provider's own page and when to pick them over us.
Get started
- Run an agent yourself with the open-source library.
- Give a task to a hosted agent and get the result back.
- Connect your own code to a cloud browser.
- Let your coding agent use your Chrome with Browser Harness.
