browz-mcp

Run the Browz MCP server from Python tooling, and drive Browz from Python. Browz keeps each web account in its own isolated browser profile; the MCP server lets an AI agent list, create, start and stop those profiles, open pages, click, type, fill forms, read the page and take screenshots in them.

The MCP server ships inside the Browz desktop app. This package finds it and starts it, so Browz must be installed and open on the same computer (download).

Use it from an MCP client

uvx browz-mcp                 # stdio, for Claude Desktop, Claude Code, Cursor, Gemini CLI, …
pipx run browz-mcp            # the same with pipx
uvx browz-mcp --http          # local Streamable HTTP on 127.0.0.1
uvx browz-mcp --where         # print the path of the server bundled with Browz

In a client's JSON config:

{ "mcpServers": { "browz": { "command": "uvx", "args": ["browz-mcp"] } } }

Claude Code: claude mcp add browz -- uvx browz-mcp.

Setup for every client (Claude, OpenAI, Grok, Gemini, Cursor, VS Code, agent frameworks) and the remote server https://mcp.gobrowz.com/mcp: https://gobrowz.com/docs/ai-agents.html

The Python SDK

pip install browz-mcp

import os
from browz import Browz

bz = Browz(token=os.environ["BROWZ_TOKEN"])   # the Local API of the Browz app on this computer
profile = bz.profiles.create(name="research")
session = bz.browser.start(profile["id"])
print(session.cdp_endpoint)                   # connect Playwright, browser-use or Stagehand here

token can also be a local agent key (blcl_…) made in Browz or with bz.agent_keys.create(); the key then limits which routes, profiles and sites the client reaches. The owner manages keys with bz.agent_keys (list, create, revoke, resume, stop_all, audit) and stored credentials with bz.credentials (list, add, update, remove). No method returns a stored password or TOTP seed.

browz.agents has session helpers for browser-use and LangChain. No dependencies; the playwright extra pins the Playwright version Browz is tested with.

mcp-name: com.gobrowz/browz