English | 中文

BrowserHand

Give any AI agent a real browser.

Every browser automation tool launches a new, blank browser — no logins, no cookies, no extensions. Your agent has to re-authenticate everywhere, fight captchas, and gets flagged as a bot.

BrowserHand takes a different approach: it controls the browser you're already using.

Why BrowserHand

	BrowserHand	Manus AI	Playwright / Puppeteer	Browser MCP
Browser	Your real browser	Your real browser	New blank browser	New blank
Login state	Preserved	Preserved	None	None
Bot detection	Much harder to detect	Much harder to detect	Often flagged	Often flagged
Open source	Yes (MIT)	No	Yes	Yes
Self-hosted	Yes	No (SaaS only)	Yes	Yes
API access	HTTP + JSON + MCP	None	Library API	MCP only
Build your own agent	Yes	No	Yes	Yes
Framework lock-in	None — curl works	Manus platform	Playwright SDK	MCP clients

Two Ways to Use

For Users — Personal Browser Automation

Install the Chrome extension, connect via MCP, and let AI control your browser directly. No coding required.

What you can do:

• Ask Claude to post on Twitter, reply to mentions, schedule threads
• Automate LinkedIn engagement — publish articles, send connection requests
• Monitor competitor websites for changes
• Fill forms, scrape data, take screenshots — all from natural language

Setup:

1. Install the BrowserHand Chrome Extension
2. Add to your Claude Desktop / Claude Code config:

{
  "mcpServers": {
    "browserhand": {
      "command": "npx",
      "args": ["browserhand-mcp"]
    }
  }
}

3. Say "Open Twitter and post: Hello World" — done.

For Agent Developers — Build Your Own Manus AI

You're building an AI agent that needs to operate a browser? BrowserHand gives your agent the same real-browser capabilities as Manus AI — open source, fully controllable, deployed on your own platform.

What you can build:

• A social media agent that auto-posts and replies across Reddit, Twitter, LinkedIn
• A customer support agent that navigates your internal tools
• A monitoring service that checks pages and alerts on changes
• Any agent that needs to interact with websites that don't have APIs

Integration via HTTP:

import requests

def cmd(action, **params):
    return requests.post("http://127.0.0.1:29981/command",
        json={"action": action, "params": params}).json()

# Your agent logic decides what to do
cmd("navigate", url="https://twitter.com")
cmd("extract")       # get all interactive elements
cmd("click", index=5)  # click "New post" button
cmd("type", index=8, text="Hello from my agent!")
cmd("click", index=12) # click "Post"

Works with any language and any agent framework — Mastra, LangChain, CrewAI, OpenAI Assistants, or your own. Just HTTP POST. Also supports MCP integration.

See the full Developer Integration Guide for local development, cloud deployment (multi-user), all 25 actions, and response format.

Quick Start

1. Load the Chrome Extension

chrome://extensions → Developer mode ON → Load unpacked → select extension/

2. Start the Relay (not needed if using MCP)

cd relay && npm install && npx tsx index.ts

3. Control your browser

curl -X POST http://127.0.0.1:29981/command \
  -H "Content-Type: application/json" \
  -d '{"action": "navigate", "params": {"url": "https://example.com"}}'

How It Works

Claude Desktop / Cursor          Your code / Agent framework
    ↓  MCP (stdio)                   ↓  HTTP POST /command
    ↓                                ↓
MCP Server (browserhand-mcp)     ────┤
    ↓  HTTP POST /command            ↓
         Relay Server (localhost:29981)
              ↓  WebSocket
         Chrome Extension
              ↓  Chrome DevTools Protocol
         Your Browser (logged in, cookies intact)

Two entry points, one core. The MCP server translates MCP tool calls to HTTP. The relay bridges HTTP to WebSocket. The extension executes commands via Chrome's debugger API.

Security

• Localhost only — relay binds to 127.0.0.1, unreachable from the network
• Origin verification — WebSocket requires an exact chrome-extension://<32-char-id> origin (a spoofed prefix is rejected); optionally pin the allowed extension id via BROWSERHAND_EXTENSION_ID
• Sensitive page blocklist — a single gate auto-refuses every page-operating action on banking & password-manager sites and on sensitive paths (/login, /signin, /password, /payment, /checkout, /billing, /account/security)
• Request guards on /command (even in local mode) — the Host header must be a loopback name (blocks DNS-rebinding), a cross-site browser Origin is rejected, and the body must be Content-Type: application/json (blocks the CSRF simple-request bypass). Non-browser clients like curl and the MCP server are unaffected as long as they send application/json.
• Optional local auth — set BROWSERHAND_REQUIRE_AUTH=1 to also require Authorization: Bearer <token> on /command in local mode. See the Integration Guide for all env vars.
• Visual transparency — agent-operated tabs are grouped and highlighted

Documentation

• Developer Integration Guide — local dev, cloud deployment, all 25 actions, response format

License

MIT