Browser use MCP,
in your real Chrome.
HailBrowser is an open-source browser use MCP server. Add one config block and Claude Code,
OpenCode, Cursor or Codex can read, click and type in the Chrome you are already signed into —
not a blank automated browser that has to log in again. When a flow is long, your agent hands
it to run_task in ultrafast mode, the fastest browser use we have measured.
Setup
One line for
Claude Code.
Install the extension first (two minutes), then add
the server. npm run setup prints every config with your real path.
claude mcp add hailbrowser --scope user -- node /path/to/hailbrowser/mcp-server/src/index.js{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"hailbrowser": {
"type": "local",
"command": ["node", "/path/to/hailbrowser/mcp-server/src/index.js"],
"enabled": true
}
}
}Cursor, Codex CLI, Claude Desktop, Windsurf and Gemini CLI are on Connect an agent.
Why this one
Your browser, not a new one.
Many browser MCP servers start a separate automated browser. That is the right choice for tests and scraping. For work that needs your accounts, it means logging in again, or not at all.
| Question | HailBrowser MCP | A separate automated browser |
|---|---|---|
| Your logins and cookies | Already there: your normal Chrome profile | A fresh profile to sign into |
| Your current tab | Left alone: agents get their own window | Not shared |
| Long flows | run_task ultrafast, 4.2s per task, runs locally | One agent round trip per click |
| What you see | Animated border, agent cursor, Stop button, every step logged | Often headless |
| Everyday use | The same extension is a Comet-style side panel | Developer tool only |
| Licence | MIT, one public repository | Varies |
Step by step
Your agent drives each click
tabs_create, read_page, click, type,
select_option, screenshot and the rest. Elements come back as refs,
so the agent never guesses coordinates. Good for debugging your own web app from Claude Code.
Delegated
Hand the whole flow to ultrafast
run_task with mode: "ultrafast" runs HailBrowser’s own decision loop:
one small request per step, the page re-checked before every input, and a stronger model
taking over if it gets stuck. Your coding agent gets back the result, not forty tool calls.
What is a browser use MCP server?
It is a Model Context Protocol server that gives an AI agent browser tools: open a tab, read the page, click, type, choose from a menu, take a screenshot. HailBrowser is a browser use MCP server for the Chrome you are already signed into, so the agent works with your real sessions instead of a fresh, logged-out browser.
How do I let Claude Code use my browser?
Install the HailBrowser extension, then run: claude mcp add hailbrowser --scope user -- node /path/to/hailbrowser/mcp-server/src/index.js. Claude Code gets 18 browser tools plus run_task, and the side panel header turns green when it is connected.
Does it work with OpenCode, Cursor and Codex?
Yes. Any MCP client works. The docs have ready config blocks for Claude Code, OpenCode, Cursor, Codex CLI, Claude Desktop, Windsurf and Gemini CLI.
Will the agent take over the tab I am using?
No. MCP calls run in a separate agent window by default, and the side panel will not drive a tab a coding agent is using. You keep browsing while your agent works.
Why is it faster than driving the browser step by step?
An agent can hand a whole flow to run_task with mode "ultrafast". HailBrowser then runs its own decision loop locally, one small decision request per step, instead of a full round trip to the coding agent for every click. In our harness that averages 4.2 seconds per task, 26 of 26 passed.
Is it safe to give an agent my logged-in browser?
The hub listens on 127.0.0.1 only and rejects web pages and unknown extensions. Page text is treated as untrusted data, agents are told never to enter passwords or payment details, and you can turn external agents off in Settings at any time.