Codex Cursor Claude CLI

Local browser automation over MCP

Drive your real Chrome from agents or the shell.

BrowseFleetMCP runs a local stdio MCP server, pairs it with a Chrome extension, and keeps sessions isolated so multiple tools can work without fighting over the same browser state.

It was inspired by BrowserMCP, but built out of not wanting the browser side to silently die in the background and leave automation hanging.

The extension lives in this repo checkout. The server can come either from the published npx -y browsefleetmcp package or from your local built dist/index.js.

One-page setup Open client configs

Transport

Standard stdio MCP

Sessions

Leased, isolated tabs

Browser

Your existing profile

System map

One server. One extension. Multiple entry points.

01 Server

Local stdio process

The root package is the actual runtime whether you launch it from Codex, Cursor, Claude Code, Claude Desktop, or the shell.

02 Extension

Repo checkout plus connected tab

No browser action works until the extension from extension-v2/ is loaded and the current tab is connected from the popup.

03 Route

Pick your launcher

Use direct CLI commands, config JSON or TOML, or the Claude Desktop bundle path depending on the client.

Most common setup failure

If the extension is loaded but no tab is attached, tool calls fail with No connected tab.

Choose your route

Setup, configs, direct CLI, or desktop bundle.

Every page below is built for a different job. Start with install if you want the end-to-end path. Jump straight to clients or CLI if the browser side is already in place.

Install

Build the server, load the extension, connect the first tab.

The full local setup path, in order, with ports and auth token notes.

Setup

Copy the exact install block for Codex, Claude Code, or another host.

One-page recipes that combine extension setup, server choice, and client wiring.

Clients

Copy the right config for Codex, Cursor, Claude, or generic stdio.

Same server entrypoint, different launcher formats and config files.

CLI

Run BrowseFleetMCP directly with ports, broker overrides, and auth.

Use this when you want the runtime without an MCP host in the middle.

Claude Desktop

Package the repo as an MCP bundle and install the extension path.

Desktop-native install route for users who do not want to hand-edit JSON.

Skill

Drop in a Codex skill that teaches the agent how BrowseFleetMCP works.

Copy the folder into ~/.codex/skills and keep the workflow close to the tool.

Operational flow

The setup has four moves. Keeping the order clean avoids most confusion.

This is not a hosted browser service. The server runs locally, the extension connects your real browser, and the client just decides how that process gets launched.

Build the root package

Compile the MCP server and CLI from the repo root.

Load the extension

Build extension-v2/ and load it in Chrome.

Choose a launcher

Use MCP config, a CLI command, or a Claude Desktop bundle.

Connect the tab

Open the popup, confirm ports, add any auth token, and attach the current tab.

Quick launch pairs

Every client path resolves to the same command shape.

Use the published package if you want the npm release. Use the local checkout if you are actively developing the repo and want the client to run your built dist/index.js.

On Windows, some clients need cmd /c in front of npx. That note lives on the Clients page so the main page stays focused on the path selection.

{
  "command": "npx",
  "args": ["-y", "browsefleetmcp"]
}

{
  "command": "node",
  "args": ["/absolute/path/to/browsefleetmcp/dist/index.js"]
}

browsefleetmcp --port 9200 --fallback-ports 9202,9204
browsefleetmcp --auth-token your-shared-token

Ready to wire it up

Start with install if you want the cleanest path. Everything else is a launcher detail.

The docs site is published from docs/ through GitHub Pages. The repo examples remain available if you prefer copying config files directly instead of transcribing snippets.

Open setup Repo examples Open skill