Live on Product Hunt Intraface launched today. Less steering, more vacations. See the launch
Start for free

What is a browser agent?

A browser agent is an AI agent that does tasks inside a web browser: it reads the page, clicks, scrolls, fills forms and moves between tabs, either as an extension in your own browser, like Claude in Chrome, or in a browser hosted for it. It acts on whatever a page shows it, so permissions and confirmations matter.

Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.

How it works

The agent sends actions to a browser, reads back the page as a tree of elements or a screenshot, and stops to ask you before risky steps.your Chrome, or abrowser hosted for ityoubrowser agentextension or hostedthe browsertabs, forms, buttonsa website1. a task, allowed sites2. navigate, click, type3. page tree, screenshotloads pages4. confirm? (buy, post)
  1. You give the agent a task and decide which sites it may use; Claude in Chrome keeps per-site permissions that you can grant or revoke at any time.

  2. The agent reads the page as a tree of elements it can act on by reference, or as a screenshot, then picks an action: navigate, click, type, scroll.

  3. The browser performs the action, the agent reads the result, and the loop continues until the task is done or it needs you.

  4. Before high-risk actions such as publishing, purchasing or sharing personal data, it stops and asks you to confirm.

  5. Everything on a page is untrusted input: text in an email or a comment can try to give the agent orders, which is what those guardrails are for.

Source: Anthropic: Piloting Claude for Chrome

Where a browser agent runs

There are two shapes, and the choice decides what the agent can reach.

The first is an extension in the browser you already use. Claude in Chrome works this way: it sees the pages you have open, uses the sessions you are signed in to, and can click buttons and fill in forms on your behalf. Anthropic began it as a pilot with 1,000 Max users, opened it to all Max subscribers in November 2025 and to Pro, Team and Enterprise plans that December. Because it rides on your own logins, Team and Enterprise admins can set site allowlists and blocklists for the whole organisation.

The second is a browser the vendor hosts. OpenAI's ChatGPT agent takes this route, driving a browser on a virtual computer OpenAI runs rather than on yours, so nothing it does touches your tabs or your sessions. For developers, Anthropic's browser use tool is the raw version of either shape: the model emits actions like navigate, read_page and left_click, and your own code runs them against a browser you operate. Nothing runs on Anthropic's side.

Browser agent versus computer use

A computer use agent works from screenshots alone and can therefore operate any application on a desktop. A browser agent trades that reach for precision. It can ask for the page's accessibility tree, in which every element carries a reference, and then click the reference rather than a guessed coordinate. That is faster, cheaper in tokens and far less likely to miss. The browser tool also understands tabs and form fields as first-class things. When a task lives entirely inside web pages, the browser agent is the right pick; the desktop agent is for everything else.

Example: checking a shared page

Suppose your team's design review lives on a shared board, a living HTML page that teammates' coding agents keep publishing to. You ask a browser agent to open the latest version, read the comments pinned to the page, and check that the new header renders on a narrow viewport. It navigates, reads the page tree to find the comment panel, resizes, takes a screenshot and reports back with what it saw. It changed nothing, so no confirmation was needed. Ask it to reply to a comment and the "post" step is where a careful agent stops and asks.

Common mistakes

  • Turning on autonomous mode first. Start with confirmations. Anthropic's own guidance is to begin with trusted sites and keep the agent away from anything financial, legal or medical.
  • Treating page text as harmless. In Anthropic's red-team run, a fake "security" email persuaded the unmitigated agent to delete the user's emails. On a set of four browser-specific attacks, mitigations cut success from 35.7% to zero, but the lesson stands: a web page is input, not instruction.
  • Giving an agent your passwords in chat. If an agent needs to sign in somewhere, type the credentials yourself rather than handing them over as text it will carry around.
  • Enabling every tool. In the browser toolset, javascript_exec and file_upload are off by default because they widen what a hostile page can make the agent do. Leave them off unless the task needs them.

FAQ

Is a browser agent safe to use on accounts I am signed in to?
Only with limits. An extension-based agent inherits your sessions, which is what makes it useful and what makes it risky. Keep it to sites you trust, leave financial, legal and medical sites out, and use the confirmation prompts rather than full autonomy for anything that sends, buys or deletes.
What is the difference between a browser agent and an agentic browser?
An agentic browser is a browser built with the agent inside it, so the agent is a feature of the product. A browser agent is the agent itself, which may live in an extension for a browser you already use or in a browser running on a vendor's machine.
Can a browser agent be tricked by a web page?
Yes. Hidden or visible text on a page can tell the agent to do something you never asked for. Anthropic measured a 23.6% attack success rate for browser use without mitigations and 11.2% with them in autonomous mode, which is why confirmations and site permissions are not optional extras.

Stop sending HTMLover Slack.

Create an account in a minute, connect your coding agent, and make your first board.