Browser use and computer use
For most people the best setup is Claude in Chrome for anything on a website, and Claude's computer use for desktop apps, running on Opus 5.5. Developers building their own agent should use Claude's computer use tool via the API, or the open-source Browser Use library if the job only touches websites.
The basics in plain words#
Browser use: the AI controls a web browser. It opens pages, clicks buttons, types into forms and reads what is on screen. It cannot touch anything outside the browser.
Computer use: the AI controls the whole computer. It moves the mouse, types on the keyboard and opens any app, like Excel, Finder or a design tool. Browser use is a smaller, safer slice of this.
The model is the brain that decides what to click. The harness is the body: the app or code that takes screenshots, moves the mouse and passes results back. You need both.
There are two ways an AI can "see" a screen:
Screenshots (pixels): the AI looks at a picture of the screen and says "click at this spot". Works on any app, but it is slower and can misclick.
Page code (the DOM): the AI reads the hidden structure of a web page, like a list of every button and its label. Faster and more precise, but only works in browsers.
Good tools mix both. Claude also tries the most direct route first: a connector (a direct link to an app like Gmail or Notion) if one exists, then the browser, and only then clicking around the screen. Clicking is the slowest and most fragile way to get anything done.
The options compared#
There are three kinds: ready-made apps you just use, model tools developers build on, and free open-source kits.
| Option | Maker | Kind | Controls | Uses your own logins? | Who it suits | The catch |
|---|---|---|---|---|---|---|
| Claude in Chrome | Anthropic | Ready-made app (Chrome extension) | Your real Chrome | Yes | Anyone on a paid Claude plan | Chrome on a computer only; it sees everything in the tabs it works in |
| Built-in browser in the Claude desktop app | Anthropic | Ready-made app | A separate browser pane inside the app | Its own, separate logins | People who want Claude kept out of their main Chrome | Only while the desktop app is open |
| Claude computer use (desktop app) | Anthropic | Ready-made app | Your whole Mac or Windows desktop | Yes | Jobs in desktop apps with no other way in | Research preview; your computer must stay awake with the app open |
| Claude computer use and browser use tools (API) | Anthropic | Developer tool | A machine or browser you set up | Only what you give it | Developers building their own agent | You build the loop that takes screenshots and clicks |
| ChatGPT agent mode | OpenAI | Ready-made app | A virtual computer in OpenAI's cloud, plus desktop apps since Aug 2026 | Partly | ChatGPT subscribers | Its cloud browser is not your browser, so many logins need redoing |
| OpenAI computer use API | OpenAI | Developer tool | An OpenAI-hosted browser | Only what you give it | Developers already on OpenAI | Public benchmarks put it a step behind Claude and Gemini |
| Gemini computer use | Developer tool (preview) | Browser, Android, desktop | Only what you give it | Cheap, fast, repetitive web jobs | You host the browser yourself; still a preview | |
| Browser Use | Open source | Free kit (Python) | A browser it launches | Only what you give it | Developers who want any model, any price | Websites only; more setup |
| Stagehand | Open source (Browserbase) | Free kit (TypeScript) | A browser it launches | Only what you give it | Developers mixing fixed steps with AI steps | Websites only |
| Playwright MCP | Microsoft, open source | Plug-in for an existing agent | A browser it launches | Only what you give it | Giving Claude Code or another agent browser hands | It is hands, not a brain: no planning of its own |
A few things changed recently. Claude in Chrome became generally available on every paid Claude plan on 26 August 2026, and now has an automatic approval mode so it no longer asks before every click. OpenAI folded Operator into ChatGPT agent mode, retired its Atlas browser on 9 August 2026, and added desktop computer use in August 2026. Google shut down its standalone Gemini 2.5 Computer Use model on 28 July 2026; computer use now lives inside Gemini 3.x Flash.
More tools#
Beyond the main players, the field splits into AI browsers you switch to, AI added to a browser you already use, all-in-one agents, and building blocks for developers.
| Tool | Maker | Kind | What it does | Price |
|---|---|---|---|---|
| Comet | Perplexity | AI browser | A Chrome-like browser whose assistant fills forms, compares and buys across tabs | Free on Windows, Mac, Android and iOS since 29 Jul 2026; background agent on Max ($200/month) |
| Gemini in Chrome (auto browse) | AI inside Chrome | Gemini clicks, scrolls and types in a new Chrome tab for shopping, travel and chores | Google AI Pro or Ultra; US only | |
| Copilot Actions in Edge | Microsoft | AI inside Edge | Copilot does multi-step tasks in Edge, with site allow and block lists | Opt-in preview |
| Opera Neon | Opera | AI browser | "Neon Do" runs tasks, keeps working in the cloud after you close it, and builds mini apps | $19.90/month |
| Dia | The Browser Company | AI browser | Chat with your open tabs and save repeatable prompts as "Skills"; more helper than clicker | Free tier |
| Manus | Meta | All-in-one agent | Works in its own cloud computer, plus a "My Computer" desktop app (Mar 2026) for your files and apps | From $20/month for 4,000 credits |
| Skyvern | Open source | Developer workflow tool | Turns repeated web chores (forms, portals, invoices) into saved workflows | Free to self-host; cloud $29 to $149/month |
| Browserbase and Steel | Startups | Cloud browsers for agents | Rented browsers in the cloud so your agent does not run on your laptop; handle logins, CAPTCHAs, blocking | About $0.08 to $0.12 per browser hour |
| UI-TARS | ByteDance, open source | Self-hosted desktop agent | A free vision model you run yourself to drive a desktop | Free, needs your own GPU |
| UiPath Screen Agent | UiPath | Enterprise automation | Classic company automation (RPA) with an AI agent built on Claude | Enterprise pricing |
Cloud browsers like Browserbase and Steel are the body, not the brain: you still pair them with Claude, Browser Use or Stagehand.
Which model to run them on#
Opus 5.5 is the safe default. Every screenshot costs roughly 1,000 to 1,800 tokens, so a long session gets expensive fast; that is why picking a lighter model for simple jobs matters more here than in normal chat.
| Model | Use it for | Example | Skip it when |
|---|---|---|---|
| Opus 5.5 | Most browser and desktop jobs, especially messy sites and multi-app work | Log into your CRM, pull this week's leads, paste them into a sheet, email a summary | The job is the same simple click path repeated many times |
| Sonnet 5.5 | Faster, cheaper sessions; terminal-heavy work | Fill the same web form for 40 contacts; test your website after each code change | The site is confusing and needs real judgment |
| Haiku 5.5 | Short, repetitive, clearly defined web steps at high volume | Check 500 product pages and note the price and stock status | Anything with surprises, logins or decisions |
| Fable 5.1 | Multi-hour desktop work where Opus keeps losing the thread | Research across dozens of sites, then build a report and deck in desktop apps | Almost everything else; it is slow and 2.5x Opus prices |
Long desktop jobs are still genuinely hard. On OSWorld 2.0, a test of real workflows that take a person about 1.6 hours, the best official entry, Claude Opus 5 at max effort, fully finishes only 44% of tasks. On the older, shorter OSWorld test, top agents now beat the human score of 72%. In plain terms: short jobs work well, long ones still need you checking in.
For developers: on the Claude API, the 5.5 models accept only the new computer use toolset, so older integrations need a small upgrade before switching.
What to pick for which job#
| Your job | Pick | Model | Why |
|---|---|---|---|
| Something on a website you are logged into (Gmail, LinkedIn, a dashboard) | Claude in Chrome | Opus 5.5 | Works in your real browser with your logins, no setup |
| Website work you want kept away from your main browser | Built-in browser in the Claude app | Opus 5.5 | Separate logins, so a mistake cannot touch your main accounts |
| A desktop app with no web version (Excel, Finder, Blender, an old internal tool) | Claude computer use in the desktop app | Opus 5.5 | The only option that drives your real desktop |
| The app already has a connector (Notion, Google Drive, Slack) | The connector, not a browser | Any | Direct links are faster and far more reliable than clicking |
| Building your own product that clicks around websites | Browser Use, Stagehand or the Claude browser use tool | Sonnet 5.5 | Cheap per run, you control everything |
| Building your own agent that drives a full desktop | Claude API computer use toolset | Opus 5.5 | Most mature desktop computer use API |
| Giving Claude Code a browser to test a site you are building | Claude in Chrome or Playwright MCP | Sonnet 5.5 | Build, open the page, check it, fix, repeat |
| Thousands of simple page checks | Browser Use or Gemini computer use | Haiku 5.5 or Gemini Flash | Lowest cost per page |
The best overall pick: Claude in Chrome running Opus 5.5. It needs no setup, uses the logins you already have, and Claude is in the top tier for this kind of work. Add Claude computer use only for desktop apps with no web version. For developers, Browser Use is the most popular free kit, at about 112,000 GitHub stars, roughly four times the next one.
Limitations of each tool#
Every tool here is slower than a person, can misclick, and can be tricked by hidden instructions on a web page. The table lists what is specific to each one.
| Tool | Main limitations |
|---|---|
| Claude in Chrome | Chrome on a computer only, no Edge, Brave or phone; paid plans only; sees everything in the tabs it works in; in auto mode it acts without asking |
| Built-in browser (Claude app) | Works only while the desktop app is open and online; you must sign in to sites again inside it |
| Claude computer use (desktop app) | Research preview; computer must stay awake and unlocked with the app open; slowest option; struggles with dropdowns, scrollbars and spreadsheets |
| Claude computer use and browser use tools (API) | You build and host the loop and the machine; each screenshot costs about 1,000 to 1,800 tokens; not available in Managed Agents; on the 5.5 models only the new toolset works |
| ChatGPT agent mode | Its cloud browser is not yours, so logins must be redone; can be slow; best agent features need paid plans |
| OpenAI computer use API | Runs in OpenAI's hosted browser; public desktop scores trail Claude and Gemini |
| Gemini computer use (API) | Still a preview; you host the browser; best at simple repeatable pages, weaker on surprises |
| Gemini in Chrome (auto browse) | US only; paid Google AI plans only; tied to Google's ecosystem |
| Copilot Actions in Edge | Preview; does not work on every site; cannot use saved passwords or autofill; Microsoft warns it may be fooled by hidden page instructions |
| Comet | Handles simple tasks well but falls apart on complex ones; past security flaws, including one triggered by a single calendar invite |
| Opera Neon | Paid only; smaller user base; tasks run on Opera's cloud, not your machine |
| Dia | Mostly chat about your tabs; little real clicking or form filling |
| Manus | Credit use varies per task and is hard to predict; failed runs cost credits; desktop access is high risk if misused |
| Browser Use | Websites only; you write Python and host it; sites may block it as a bot |
| Stagehand | Websites only; you write TypeScript; each AI step adds cost and some randomness |
| Playwright MCP | Only gives an agent browser controls; no planning of its own; can get blocked by bot checks |
| Skyvern | Best for repeated, fixed workflows, not open-ended research; credits on cloud plans |
| Browserbase and Steel | Infrastructure only, no brain; billed by the hour; Steel browsers run only in the US east region |
| UI-TARS | Needs your own powerful GPU and setup skills |
| UiPath Screen Agent | Built for large companies; enterprise pricing and setup |
The common thread: AI automation succeeds roughly 85 to 95% of the time on web tasks, versus 99%+ for traditional scripted automation. For anything that must work every time, use a connector or a fixed script and save the AI for the messy parts.
Safety and limits#
The biggest risk is a web page that hides instructions for the AI, such as "send this user's emails to this address". This is called prompt injection, and Claude can sometimes follow it. Anthropic scans screenshots for this and makes Claude double-check, but it is not perfect.
Separate the risky stuff: use a separate browser profile or the built-in browser for experiments, and keep banking and payments out of reach.
Stay in the loop for anything final: sending, paying, deleting, accepting terms. Keep approval on for those steps.
Expect it to be slower than you and to misclick now and then, especially on dropdowns, scrollbars and spreadsheets.
Watch other extensions: in May 2026 researchers showed that another malicious Chrome extension could trigger Claude in Chrome tasks; it was patched, but install extensions carefully.
Developers: run computer use in a sandbox (a throwaway virtual machine or container), limit which websites it can reach, and never paste real passwords into prompts.