Browser use,
in your browser.
Browser use is an agent doing the clicking for you: reading a page, filling a field, choosing a result, checking it worked. HailBrowser does that inside the Chrome you already use, at a 4.2s average over 26 tasks — and it exposes the same browser to Claude Code, OpenCode and Cursor over MCP. If you came looking for MCP browser use, that is this: one config block, then your agent drives the browser you are already signed into.
The category
An agent with hands.
A chat model can tell you what to click. Browser use is the part where something actually clicks it, sees the result and carries on. That needs a browser, a way to read the page, and a loop that decides the next action.
- Read
- Turn the page into something a model can reason about: elements, labels, values, not pixels.
- Act
- Click, type, choose, press a key, scroll, open a tab, accept or dismiss a dialog.
- Verify
- Look at what changed. A field that rejected a value is a failure, not a step completed.
- Report
- Say what happened, where it stopped, and what needs a human. No invented success.
Two ways in
Browser use, or browser use over MCP.
Most people want the first one: type a job, watch it happen. Developers want browser use MCP: their own agent, using this browser as a tool. Both are the same extension.
01 · Side panel
The everyday way
Click the HailBrowser icon and ask in plain English. It opens the tab, reads the page, fills the form and writes the answer with sources. Every step is listed, with a Stop button.
Chat for a conversation, Fast for forms and searches, Research for cited answers, Tasks for long jobs. One browser, four moods.
Read the modes02 · MCP
The agent way
HailBrowser runs a local MCP server with 18 browser tools on 127.0.0.1:12306.
Add one config block to Claude Code, OpenCode, Cursor or Codex and your agent can drive the
same signed-in browser.
- Claude Code
- OpenCode
- Cursor
- Codex CLI
- Claude Desktop
- Gemini CLI
Compared
Browser use, two shapes.
browser-use is a strong open-source Python library that drives a browser it launches. HailBrowser is an extension that drives the browser you are in. Both are legitimate; the question is which one fits your day.
| Question | HailBrowser | A Python browser-use library |
|---|---|---|
| Which browser? | The Chrome profile you already use | A Chromium the script launches |
| Your logins | Already there | Sign in again inside that browser |
| How you drive it | Side panel, or MCP from your agent | A Python script or your own service |
| Setup | Load the extension, add a model key, optionally add one MCP config block | Install the package, write the harness, provide a model |
| Agent integration | Any MCP client | Python API |
| Speed evidence | 26 tasks at 4.2s, harness published | Measure your own task set |
| Source and licence | MIT, extension and server in one repo | Check the project’s repository |
Use the Python library when the browser should be disposable and the harness should live in your code. Use HailBrowser when the browser is the one you are already signed into, or when your agent already speaks MCP.
Also an AI browser
If what you want is a Comet-style assistant in the browser you already have, that is the same extension: a side panel that reads the page and does the work. The open-source Comet alternative page compares the two honestly.
What is browser use?
Browser use is the category of AI agents that operate a web browser to finish a task: reading a page, clicking, typing, choosing from a menu and checking the result. It covers Python frameworks that drive their own browser, and agents built into a browser, like HailBrowser.
Does browser use work with my logins?
With HailBrowser, yes. It drives your normal Chrome profile through the extension, so the sites you are already signed into work without copying cookies or starting a second profile.
Can Claude Code use my browser through MCP?
Yes. HailBrowser runs a local MCP server on 127.0.0.1:12306 and exposes 18 browser tools. Claude Code, OpenCode, Cursor, Codex, Windsurf, Gemini CLI and Claude Desktop can each drive the same browser after one config block.
How fast is it?
In our harness, Ultrafast passed 26 of 26 tasks at a 4.2 second average across three full runs, with no false completions. The harness and every per-task number are published on the benchmarks page.