Use browser use for work that can finish inside a browser session. Use computer use when an agent has to reach a native app, desktop window or system dialog. The distinction is the area the tool can act in. A browser tool can click by coordinates and a desktop tool can operate a browser window, so the presence of a mouse tells you little about the permission you are granting.
One Reddit user described getting blocked on signed-in sites and seeing longer runs fail partway through. That is a useful reason to test your own session with a bounded task.
There is a third option when only one step leaves the browser. Let a browser agent extract or download the source, then open the result yourself or pass it to a dedicated file tool. This avoids granting a whole desktop session for a short handoff. Tabbit Browser is one example of a browser where the browsing and delegation steps can happen together.
Key takeaways
Start with the smallest environment that can produce the requested result.
Separate where a tool can act from how it locates a control. A browser can provide both structured elements and screenshots.
A browser-only tool can still act as a signed-in user. Check which profile or remote session it will use.
A screenshot records what was visible at one moment. Capture a fresh one after a resize, navigation or dialog.
Verify the final page state or file yourself before relying on a delegated run.
Browser and desktop tools at a glance
| Decision | Browser use | Computer use |
|---|---|---|
| Operating area | One browser session, page or group of tabs | Graphical environment exposed to the tool |
| What the tool can inspect | Page elements, accessible names, text, screenshots, depending on interface | Screenshots, sometimes accessibility information or window controls |
| Typical actions | Navigate, click, fill, extract, switch tabs | Move pointer, type, scroll, switch applications |
| Permission to check | Browser profile, origin, tabs and authentication | Desktop session, apps, local files and operating-system permissions |
| Example | Open public policy details and copy a clause | Open a downloaded file in a native viewer and use its Print dialog |
Which interface can see the control?
A browser action can be semantic. A tool can ask for a button by its accessible role and name, then look it up again just before acting. This works for a normal button with a unique label. A page can also paint a chart or editor on a canvas, where the controls the user sees may have no useful semantic targets. A screenshot and a browser-scoped coordinate click may solve that case while staying in the current page. Playwright locators and the browser automation guide describe the structured path.
A graphical computer tool can see whatever appears in its allowed display area. A model reads a screenshot, returns an action such as moving the pointer and clicking, and receives an updated view. A desktop file picker or a standalone PDF viewer is outside the web page, so browser locators cannot discover its buttons. Before escalating, check whether the browser tool can upload a known file through an input, or whether a file-processing tool can open the result directly. A native dialog may be avoidable even when it appears in the manual workflow. See the Anthropic computer-use documentation.
Suppose you need to read a cancellation clause behind a button. If the page exposes a button with a stable accessible name, look it up by role and name. When the viewport narrows and the button moves, Playwright resolves the locator again before the next action. If the interface only exposes pixels, get a fresh view and locate the button again before clicking. In either case, verify the revealed clause and its source URL.
This is a choice about input, separate from where the tool runs. Anthropic’s browser tool documents both page references and viewport coordinates. For a recurring job that reads the same field, check for a supported export or API first. If the site offers neither, prefer a stable page locator and verify the output before retrying. A retry after an uncertain form submission can create a duplicate record. Anthropic browser-use documentation, browser-agent controls.
Where the browser ends
Consider a workflow that reads a warranty page, downloads a PDF and sends only page two to a native print dialog. There are three different results to verify. The warranty clause should match the source URL. The downloaded file should be the right document. The print job should target the correct pages and device. The first two may fit a browser and a file tool. The last step requires a tool that can reach the print UI, or a person who can confirm those settings.
Ask exactly what the external step is before broadening access. If the real request is “extract page two,” direct file processing may be enough. If the request is “use these printer settings,” have the desktop tool stop at the review screen. It is harder to recover once a job is sent to a shared device. The desktop assistant guide explains what that additional environment covers.
Protect the session you actually grant
Read-only intent does not make a signed-in browser read-only. A session that can see invoices may also be able to open account settings or change a billing preference. Start with public URLs or a separate profile when possible. Limit the first run to named domains. Define what to do if a login, permission prompt, account change or unfamiliar download appears. See the browser agent security guide and the virtual browser guide.
A Codex user reported using browser control to navigate a complex UI for testing and debugging. That report is specific to their experience.
Desktop access can expose unrelated windows and files. Give the agent a separate desktop or virtual environment where possible, close unrelated apps and use only the files needed for the task. Treat text on a website or in a document as input to inspect. A page can ask the agent to ignore your instruction or follow a link, and it cannot grant that permission.
Try a browser-only handoff in Tabbit
For the warranty example, start with the source page and provide the exact field to extract. On the international Tabbit site, the product shows how to attach a tab or screenshot as context and have its Agent plan and run a browser task. If you only want a clause explained, use page chat. If you want the agent to open the policy and return a cited answer, give it a bounded request in Agent Mode.
Read the policy linked from this public page. Return the exact cancellation
window, a short excerpt and the source URL. If a sign-in is required, stop
and report which page asked for it. Do not submit anything or download files.This handoff keeps the evidence in the browser. For a local print or export step, hand the result to a person or a tool designed for it. For the product's context and execution controls, see how Tabbit works.
Choose by the result you need
| You need | First tool | Success check |
|---|---|---|
| Compare public policy pages | Browser tool | Source URL and exact wording per row |
| Use a site editor with a canvas | Browser visual action, if available | Changed value or canvas state |
| Pull text from a local PDF | File tool | Correct page and text |
| Operate a native print dialog | Desktop tool or person | Correct printer, range and approval before sending |
| Review a workflow that spans both | Split at the app boundary | Evidence for each handoff and final result |
The browser and desktop guides for Terra browser use and Terra computer use explain why model capability and tool permission are separate. For an initial run, choose a task whose end state you can check in a few seconds. Add a wider environment only when you can name the action that needs it.
Tabbit fits the first step when you need to gather and check evidence from web pages in the same browser.
FAQ
What is the difference between browser use and computer use?
Browser use operates a browser or a browser session. Computer use operates a graphical environment and can extend to desktop apps and operating-system dialogs if the tool has access to them.
Does computer use always need screenshots and a mouse?
A screenshot-driven tool reads screenshots and acts through pointer and keyboard input. Some desktop automation also uses accessibility trees or app APIs, so check the actual interface you are being asked to grant.
Can a browser agent work on a website where I am already logged in?
Some agents can connect to a supported browser profile, while others launch a clean or remote session. Confirm which session the agent can see and what account access it inherits.
Can a browser tool click by coordinates?
Yes. Browser automation libraries can send pointer input to a web page by coordinates. The tool still operates within its browser scope.
When do I need a desktop tool instead of a browser tool?
Use a desktop tool when the result requires interaction with a native app or system dialog and there is no suitable API or file handoff. A workflow can use a browser tool first and switch only for that step.
How can I check whether the agent really finished?
Look for the state change you requested, such as visible terms, a verified source link, or a saved file in the correct location. A click or a status message only confirms that a step was attempted.