Airbrowser

Open-source browser automation API with anti-detection — Undetectable Chrome for AI agents, web scraping, and automation. REST API + MCP server + VNC debugging. Selenium/Playwright alternative that bypasses Cloudflare.

Quick Start

Cloud Hosted (no setup)

Use the managed cloud version - no installation required:

https://airbrowser.dev

Docker (one-liner)

docker run -d -p 18080:18080 --name airbrowser ghcr.io/ifokeev/airbrowser-mcp:latest

# With NVIDIA GPU (recommended for anti-detection)
docker run -d -p 18080:18080 --gpus all --device /dev/dri:/dev/dri --name airbrowser ghcr.io/ifokeev/airbrowser-mcp:latest

Portable Downloads

Download and run - no Docker knowledge required:

Platform	Download	Requirements
Linux	airbrowser-linux.tar.gz	uidmap package or Docker
macOS	airbrowser-mac.tar.gz	Colima, Docker Desktop, or Podman
Windows	airbrowser-windows.zip	Docker Desktop or Podman

# Linux/macOS
tar -xzf airbrowser-*.tar.gz && cd airbrowser-* && ./airbrowser

# Windows: Extract zip and double-click airbrowser.bat

From Source

git clone https://github.com/ifokeev/airbrowser-mcp.git
cd airbrowser-mcp
docker compose up --build

# With NVIDIA GPU
docker compose -f compose.gpu.yml up --build

Local Mode (no Docker) — Linux only

Run natively without a container — zero container fingerprint for maximum anti-detection stealth. Tested on Ubuntu/Debian.

git clone https://github.com/ifokeev/airbrowser-mcp.git
cd airbrowser-mcp
uv run python run_local.py         # auto-installs deps + system packages
uv run python run_local.py --vnc   # with VNC viewer at http://localhost:6080/vnc.html

Requires Chrome installed on the host. See python run_local.py --help for options.

Service	URL	Description
Dashboard	http://localhost:8000/dashboard	Browser pool management UI
Swagger Docs	http://localhost:8000/docs/	Interactive API documentation
REST API	http://localhost:8000/api/v1/	Browser automation endpoints
MCP Server	http://localhost:3099/mcp	Model Context Protocol for AI agents
VNC	vnc://localhost:5900	Remote desktop (with --vnc flag)
noVNC	http://localhost:6080/vnc.html	Web-based VNC viewer (with --vnc flag)

---

Open http://localhost:18080 - all services available:

Service	Path
Dashboard	/
API Docs	/docs/
REST API	/api/v1/
MCP Server	/mcp
VNC Viewer	/vnc/

Features

• Undetected Chrome (SeleniumBase UC)
• 100+ concurrent browsers
• Persistent profiles & cookies
• Tab management
• Proxy per browser (DataImpulse recommended)
• MCP for AI agents
• AI vision tools (optional)

GPU Passthrough (Recommended)

GPU passthrough enables hardware-accelerated WebGL rendering via Vulkan, making the browser fingerprint match a real desktop machine. Without it, Chrome falls back to software rendering (SwiftShader) which is easily detected by anti-bot systems.

Requirements: NVIDIA GPU + NVIDIA Container Toolkit

# Docker Compose (recommended)
docker compose -f compose.gpu.yml up

# Docker run
docker run -d -p 18080:18080 \
  --gpus all \
  --device /dev/dri:/dev/dri \
  -e NVIDIA_VISIBLE_DEVICES=all \
  -e NVIDIA_DRIVER_CAPABILITIES=all \
  ghcr.io/ifokeev/airbrowser-mcp:latest

# Portable launcher
./airbrowser --gpu

Without a GPU, Chrome uses --use-gl=swiftshader automatically. With GPU passthrough, it uses --use-gl=angle --use-angle=vulkan for real GPU rendering.

AI Vision (Optional)

Enable AI-powered vision tools (what_is_visible, detect_coordinates) with any OpenAI-compatible vision backend. Vision turns on only when VISION_API_BASE_URL, VISION_API_KEY, and VISION_MODEL are all set.

When smart targeting is enabled per request, detect_coordinates can validate a raw vision point, optionally snap to a nearby clickable target, and return both the original click_point and a resolved_click_point with an outcome_status that tells you whether the result was confirmed, corrected, or needs inspection before clicking. Pair that with gui_click or MCP-compatible gui_click_xy to re-check coordinate clicks and request post-click feedback.

# Docker run
docker run -d -p 18080:18080 \
  -e VISION_API_BASE_URL=https://your-openai-compatible-endpoint/v1 \
  -e VISION_API_KEY=your-api-key \
  -e VISION_MODEL=your-vision-model \
  ghcr.io/ifokeev/airbrowser-mcp:latest

# Docker compose
VISION_API_BASE_URL=https://your-openai-compatible-endpoint/v1 \
VISION_API_KEY=your-api-key \
VISION_MODEL=your-vision-model \
docker compose up

MCP Client Configuration

Add airbrowser to your AI coding assistant:

<details> <summary><b>Claude Code</b></summary>

claude mcp add airbrowser --transport http http://localhost:18080/mcp

</details>

<details> <summary><b>Cursor</b></summary>

Go to Cursor Settings → MCP → Add new MCP Server:

{
  "mcpServers": {
    "airbrowser": {
      "url": "http://localhost:18080/mcp",
      "transport": "http"
    }
  }
}

</details>

<details> <summary><b>VS Code / Copilot</b></summary>

Add to your MCP settings:

{
  "mcpServers": {
    "airbrowser": {
      "url": "http://localhost:18080/mcp",
      "transport": "http"
    }
  }
}

</details>

<details> <summary><b>Cline</b></summary>

Follow Cline MCP guide with:

{
  "mcpServers": {
    "airbrowser": {
      "url": "http://localhost:18080/mcp",
      "transport": "http"
    }
  }
}

</details>

<details> <summary><b>Windsurf</b></summary>

Follow the Windsurf MCP guide with the config above. </details>

Test your setup

Navigate to https://example.com and take a screenshot

Your AI assistant should create a browser, navigate to the URL, and return a screenshot.

Generated Clients

Auto-generated from OpenAPI spec:

# Python
pip install airbrowser-client

# TypeScript
npm install airbrowser-client

Community

Join our Discord server for support, feature requests, and discussion.

Docs

• docs/ - Full documentation
• examples/ - Code samples

License

Fair Source - Free for up to 10 users. Cannot be offered as a hosted service. Commercial license required for larger deployments.