Skip to content

Browser MCP 2.0 is here. Upgrading from 1.x? Update both the extension and the MCP server. How to update

LogoBrowser MCP

Let AI use your real browser

Browser MCP connects Claude Code, Cursor, Codex and other MCP apps to Chrome. Agents work in tabs of their own, signed in where you are, and you can see and stop them at any time.

Claude Code
https://tables.example.com/nonnas
Logo
“Browser MCP” started debugging this browserCancel
Claude Code

Book a table for 4 at Nonna's this Friday, around 7:30.

A replay of real browser tool calls. Pick another tab to see another task.

Your browser, not a fresh one

A headless browser starts every task signed out, and can't see the tab you're looking at. Browser MCP works in the Chrome you already have open.

Already signed in
Agents use your browser profile, so they reach the dashboards, inboxes and admin pages you can, without you handing over a password.
Beside you, not instead of you
Agents open their tabs in the background, in a group named after the AI app. The rest of the browser stays yours.
Hand over a tab
Halfway through a form or a checkout? Drag the tab into the agent's group, or tell it which tab to use. Drag it out to take it back.
Fewer calls, fewer tokens
A whole task often fits in one tool call. Each action answers with one line instead of a page snapshot, and agents read only what they ask for.

One tool, plain commands

Agents get a single browser tool that takes command lines. Targets are Playwright locators, CSS selectors or refs from a snapshot, so most steps need no snapshot first. Every line is checked before any of them runs.

Full command reference

Go places

  • goto <url>
  • back forward reload
  • tab list
  • tab select <tabId>
  • tab new [<url>]

Act

  • click <target>
  • fill <target> "<text>"
  • press <key>
  • select <target> "<value>"
  • check <target>
  • upload <target> <path>
  • dialog accept

Read

  • snapshot -i
  • find "<text>"
  • get text <target>
  • screenshot --full

Wait

  • wait --text "<text>"
  • wait --url <glob>
  • wait <target>

Debug

  • console --errors
  • errors
  • requests --filter "<text>"
  • eval "<function>"

And with --caps cdp, anything the Chrome DevTools Protocol can do: breakpoints, response mocking, profiling, emulation.

Set up in three steps

You need Chrome or another Chromium browser such as Edge or Brave, Node.js 20 or later, and an AI app that supports MCP.

  1. Add the extension

    Then pin it to the toolbar, so you can see what agents are doing.

  2. Add the server to your AI app

    Then restart the app. It starts the server with npx when it needs it.

    claude mcp add --scope user browsermcp -- npx @browsermcp/mcp@latest
    Every app's steps, and options
  3. Ask, then allow once

    The first time, the agent answers with a six-digit code. Check that the extension's popup shows the same code, click Allow, and ask again. From then on, agents use the browser without asking.

    Your agent

    Click Allow in the Browser MCP extension's popup. The code is 482 913

    The extension's popup

Coming from Browser MCP 1.x? Update both parts

The extension and the MCP server are both 2.0 now, and only work together. If Browser MCP says one of them is outdated, update both.

See everything that's new
The extension
Chrome updates it within a day. To update now, open chrome://extensions, turn on Developer mode and click Update. Chrome then asks you to accept a new permission ("View and manage your tab groups"); until you do, the extension is turned off.
The MCP server
Use @browsermcp/mcp@latest in your AI app's MCP config (replace a pinned version such as @0.1.3) and restart the AI app.
What's new
Agents no longer need Connect on each tab: you click Allow in the popup once, and then they work in their own tabs, grouped and colored per AI app. Stop in the popup pauses them all.

You can see it. You can stop it.

Agents work in tabs you can see, only after you've allowed them once, and nothing listens beyond your own computer.

Local only
The background process listens on 127.0.0.1, so other machines can't connect. Browsing data stays on your computer, apart from what agents read and send to their AI app.
Paired with a code
The extension obeys only the server you allowed. Another program that takes over the port gets nothing, and only AI apps you run hold the token to send commands.
Off limits
Incognito tabs (unless you allow the extension there), browser pages such as chrome://settings, and the Chrome Web Store. Agents see your other tabs' titles only when they list them.
One button to stop
Stop in the popup pauses every agent and releases its tabs. Cancel on Chrome's debugging bar does the same.

Browser MCP protects your browser from web pages and from other programs that take over its port. It can't protect it from malware running as your user: only your operating system can. How it works

How a command reaches your tabs
  1. Your AI app

    calls the browser tool

    browser

  2. MCP server

    one per AI app

    @browsermcp/mcp

  3. Background process

    one per user, on 127.0.0.1

    127.0.0.1:9010

  4. Extension

    debugs only agents' tabs

    chrome.debugger

  5. Your tabs

    in a named group per app

Give your agent a browser

Add the extension, add the server to your AI app, and ask for something you'd normally click through yourself.