agnt 0.5.1

MCP server for AI coding agents - process management, reverse proxy with traffic logging, browser instrumentation, and sketch mode

agnt

MCP server for AI coding agents - process management, reverse proxy with traffic logging, browser instrumentation, and sketch mode.

Installation

pip install agnt
# or
uv pip install agnt

Quick Start

As MCP Server (Claude Code, etc.)

Add to your Claude Code MCP configuration:

{
  "mcpServers": {
    "agnt": {
      "command": "agnt",
      "args": ["serve"]
    }
  }
}

Or install as a Claude Code plugin:

/plugin marketplace add standardbeagle/agnt
/plugin install agnt@agnt

As PTY Wrapper

Wrap your AI coding tool with overlay features:

agnt run claude --dangerously-skip-permissions
agnt run gemini
agnt run copilot

Features

• Project Detection: Auto-detect Go, Node.js, Python projects
• Process Management: Run and manage long-running processes
• Reverse Proxy: HTTP proxy with traffic logging
• Browser Instrumentation: 50+ diagnostic primitives
• Sketch Mode: Excalidraw-like wireframing on your UI
• Floating Indicator: Quick access panel in browser

MCP Tools

Tool	Description
detect	Detect project type and available scripts
run	Run scripts or commands
proc	Manage processes: status, output, stop, list
proxy	Reverse proxy: start, stop, exec, toast
proxylog	Query proxy traffic logs
currentpage	View active page sessions
daemon	Manage background daemon

Usage Examples

# Start a proxy for your dev server
proxy {action: "start", id: "dev", target_url: "http://localhost:3000"}

# Execute JavaScript in connected browsers
proxy {action: "exec", id: "dev", code: "__devtool.screenshot('homepage')"}

# Show toast notification
proxy {action: "toast", id: "dev", toast_message: "Build complete!", toast_type: "success"}

Browser API

The proxy injects window.__devtool with 50+ functions:

• screenshot(name) - Capture screenshot
• inspect(selector) - Get element info
• sketch.open() / sketch.save() - Wireframe mode
• indicator.toggle() - Toggle floating indicator
• And many more...

Configuration

Create .agnt.kdl in your project root:

scripts {
    dev {
        command "npm"
        args "run" "dev"
        autostart true
    }
}

proxies {
    frontend {
        target "http://localhost:3000"
        autostart true
    }
}

Documentation

• GitHub
• Documentation

License

MIT