⚡ Standalone Native Rust Engine · 43 Tools · Zero Headless Lag

Next-Generation Browser Control
for AI Assistants

Give Claude Code, Cursor, Gemini/Antigravity, Windsurf, and VS Code direct control of your live, authenticated Chrome and Brave browser sessions. 43 tools, sub-4ms cold startup, ~3.8MB RAM, and zero headless login barriers.

< 3.8 ms
Cold Boot Latency (Native Rust)
~3.8 MB
RAM Footprint (95% cut vs Node)
43 Tools
Full CDP & Everyday Tool Surface
20 Sessions
Color-Coded Tab Group Isolation
Live execution: Claude Code driving an authenticated browser session in real time with sub-4ms CDP event dispatch.
Active Tab Bridge

Instant Client Setup

Connect your assistant directly to your daily browser in 60 seconds with copy-paste configs.

Config: ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "browser-control": {
      "command": "browser-mcp"
    }
  }
}
Load the unpacked extension in Chrome/Brave Developer Mode from ~/.browser-mcp/extension (or run browser-mcp install to unpack automatically).

Interactive Architecture

Click any layer to inspect protocol guarantees, security boundaries, and telemetry specifications.

Layer 1

AI Coding Assistant

Claude Code, Cursor, Gemini, Windsurf, or VS Code client.

stdio JSON-RPC 2.0
Layer 2

Native Rust Core

Standalone binary engine with sub-4ms boot and ~3.8MB RAM.

Port 9876-9895
Layer 3

Autonomous Token Shield

Semantic DOM pruning, monotonic [e1] handles, and local spooling.

Origin Verified
Layer 4

Live Chrome & Brave

Authenticated user tabs, CDP compositor overlays, and multi-session isolation.

Persistent MV2
Layer 1: AI Coding Assistant & stdio Interface
Your coding assistant (Claude Code, Cursor, Gemini, Windsurf, or custom agent) invokes standard Model Context Protocol methods over standard input/output. Asynchronous reader and writer tasks process requests with zero external daemons.
Protocol Interface JSON-RPC 2.0 (stdio)
Client Dispatch < 0.5 ms event-loop
Process Model Zero persistent daemon

Real-World Automation Scenarios

Inspect how Browser Control handles both everyday user tasks and advanced developer workflows.

browser-control-session.log
Tokens: 1,420 (98.6% pruned) · Latency: 2.4ms
https://store.enterprise.com/checkout
● Session Isolated Tab #142 (Tab Group: Blue)
Shipping Address & Billing
12 fields batched synchronously via browser_fill. Human-in-the-loop modal active for SMS OTP confirmation.
[CDP Compositor Overlay] Highlight Quad active on #btn-verify-otp

Interactive Tool Directory

Explore all 43 native browser tools covering both everyday user automation and high-intensity developer workflows.

Showing 43 of 43 tools
Open Full Tools Documentation →

Autonomous Token Shield

Eliminate LLM context bloat and rate-limit penalties with intelligent local pruning.

Semantic & Viewport DOM Pruning

Raw webpage DOMs routinely surpass 300,000 tokens. Token Shield condenses pages into an accessible semantic tree, pruning off-screen nodes and mapping elements to monotonic handles ([e1]-[e43]).

  • 98.8% Token Savings: 142k tokens distilled to 1.6k
  • Spatial Viewport Pruning: viewport_only=true saves 50–70%
  • Handle Generation: Instant targeting with [e1] refs

Atomic Form Batching & Spooling

Rather than burning one LLM turn per input field, Browser Control dispatches entire forms atomically with event-loop pacing, with automatic disk spooling for massive payloads.

  • Single-Turn Batching: Fill 20+ fields in one request
  • Event-Loop Pacing: 35ms debounced React/Vue events
  • Disk Spooling: Payloads >2,500 tokens stored locally

Production Proven in Real Browsers

No dummy simulations. See how Browser Control operates inside live browser tabs.

Multi-session tab groups

Multi-Session Tab Groups

Color-coded tab groups isolate independent AI agent conversations without profile locks.

Human-in-the-loop overlay

Human-in-the-Loop Safeguards

Sensitive 2FA prompts, SMS OTPs, and CAPTCHAs pause cleanly for human completion.

43 tools available

43 Automation Tools

Hardware clicks, atomic form batching, cookie extraction, and failed network telemetry.

Engineered Without Compromise

How Browser Control compares to legacy Node.js bridges and headless runners.

Architectural comparison across active browser MCP implementations.
Capability Browser Control (Native Rust) Agent360 (Legacy Node) Playwright MCP BrowserMCP.io
Runtime Engine Standalone Native Rust (0 dependencies) Node.js runtime (npm dependency) Node.js + Headless Chromium Node.js runtime (npm dependency)
Startup Latency < 3.8 ms (instant native binary) ~450 ms (Node cold start) ~2,500 ms (Chromium boot) ~500 ms
Memory Footprint ~3.8 MB RAM ~85 MB RAM ~350 MB+ RAM ~90 MB RAM
Context Token Shield Built-in (spooling >2.5k tokens + handles) None (dumps raw DOM) None (35,000+ raw tokens) None
Spatial Viewport Pruning Yes (viewport_only=true, 50-70% savings) No No No
Atomic Form Batching Yes (single-turn multi-field fill) No (1 field per turn) No (sequential actions) No
Authenticated Session Reuse Yes (active Brave/Chrome tabs) Yes (Chrome tabs) No (isolated profile default) Yes (Chrome tabs)
Diagnostics & Overlays browser_highlight + failed requests Screenshot only Trace viewer file Screenshot only
Tool count 43 41 73 Not published
Multi-Session Concurrency 20 concurrent sessions 20 concurrent sessions Single session (profile lock) Single session
Telemetry & License MIT · Zero Telemetry MIT · Zero Telemetry Apache-2.0 · Microsoft Telemetry MIT · Unmaintained

Frequently Asked Questions

Everything you need to know about architecture, privacy, and client support.

What is Browser Control and why was it rewritten in Rust?
Browser Control (formerly Browser MCP) is an open-source Model Context Protocol server that bridges AI assistants to your daily browser. It was rewritten from scratch as a standalone native Rust binary to eliminate Node.js dependencies, achieve sub-4ms cold startup, reduce idle memory to ~3.8MB, and introduce an autonomous Token Shield that protects LLM context windows from token exhaustion.
How does Browser Control differ from Playwright MCP and headless runners?
Unlike Playwright MCP, which spins up isolated headless profiles that hit bot verification and login walls, Browser Control connects directly to the Brave and Chrome tabs you are already signed into. It inherits your cookies, active logins, and 2FA authentication, allowing your AI assistant to act on private dashboards and SaaS tools without mock credentials.
Is my browsing data and token security protected?
Yes. Browser Control contains zero telemetry, zero analytics, and connects strictly over local loopback (127.0.0.1:9876-9895). It enforces strict Origin: chrome-extension://[a-p]{32} verification on incoming WebSockets to prevent unauthorized local process connections. Source code is completely open and auditable on GitHub.
How does 2FA assistance and CAPTCHA solving work?
When an automated task encounters a CAPTCHA or 2FA verification challenge, Browser Control invokes browser_ask_user to render an interactive in-browser dialog. The automation pauses cleanly until you approve or complete the step on your own screen, after which the agent resumes in the exact same tab.
Which AI coding assistants are supported?
Any assistant supporting standard MCP JSON-RPC over stdio works out of the box, including Claude Desktop, Cursor, Gemini/Antigravity, Windsurf, VS Code (Cline/Roo), Claude Code, OpenCode, and Zed. Run browser-mcp install to automatically configure all detected client configuration files.
Who authors and maintains this project?
Browser Control is engineered and maintained by Mohamed Elsheikh (@modded-soldier-9) under the permissive MIT License. Contributions and issues are welcome on GitHub.

Take Full Control of Your Real Browser

Install Browser Control in 60 seconds and give your AI assistant access to authenticated web sessions.

Install Browser Control View on GitHub