puppet-ai

Full computer control for AI agents — see, click, type, scroll via MCP.

Give any AI agent eyes and hands. puppet-ai captures the screen, reads text via native OCR, detects UI elements, and controls mouse + keyboard. Works with any app on macOS — browsers, desktop apps, games, terminals.

Why puppet-ai?

• Fast — native macOS OCR in ~0.5s, 7x faster with caching

• Universal — works with ANY app, not just browsers

• Agent-agnostic — MCP standard, plug into any AI agent

• Secure — auto-masks API keys, passwords, credit cards, emails in OCR output

• Complete — 27 tools: vision + actions + system

• Private — all processing on-device, no data leaves your Mac

Related MCP server: mcp-desktop

Quick Start

1. Install

pip install puppet-agent

2. Enable Accessibility

System Settings → Privacy & Security → Accessibility → enable your terminal/IDE app.

3. Connect to your AI agent

puppet-ai is an MCP server. Connect it to any MCP-compatible agent below, then ask the agent to interact with your computer.

Integrations

Claude Code

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "puppet-ai": {
      "command": "puppet-ai",
      "args": ["serve"]
    }
  }
}

OpenAI Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.puppet-ai]
command = "puppet-ai"
args = ["serve"]

Or via CLI:

codex mcp add puppet-ai -- puppet-ai serve

Google Gemini CLI

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "puppet-ai": {
      "command": "puppet-ai",
      "args": ["serve"]
    }
  }
}

Verify: launch gemini and run /mcp to check connection.

Google Antigravity

Via MCP settings in Antigravity, or add to your project's .antigravity/settings.json:

{
  "mcpServers": {
    "puppet-ai": {
      "command": "puppet-ai",
      "args": ["serve"]
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "puppet-ai": {
      "command": "puppet-ai",
      "args": ["serve"]
    }
  }
}

Or: Cursor Settings → Tools & MCP → Add Server.

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "puppet-ai": {
      "command": "puppet-ai",
      "args": ["serve"]
    }
  }
}

Cline (VS Code)

In VS Code, open Cline settings → MCP Servers → Add:

{
  "puppet-ai": {
    "command": "puppet-ai",
    "args": ["serve"]
  }
}

Zed

Add to Zed settings (~/.config/zed/settings.json):

{
  "context_servers": {
    "puppet-ai": {
      "command": {
        "path": "puppet-ai",
        "args": ["serve"]
      }
    }
  }
}

OpenClaw

Add to your agent's MCP config:

mcp_servers:
  puppet-ai:
    command: puppet-ai
    args: [serve]

Any MCP Client

puppet-ai speaks MCP over stdio. Spawn it as a subprocess:

import subprocess
proc = subprocess.Popen(
    ["puppet-ai", "serve"],
    stdin=subprocess.PIPE,
    stdout=subprocess.PIPE,
)
# Communicate via MCP JSON-RPC over stdin/stdout

Works with any agent that supports the Model Context Protocol.

Detailed setup guides: integrations/

Tools

Vision (see the screen)

Actions (control the computer)

System

How It Works

AI Agent (Claude, Codex, Gemini, Cursor, Windsurf, ...)
    ↕ MCP protocol (stdio)
puppet-ai server
    ├── Vision: Apple Vision OCR + CGWindowList capture
    ├── Accessibility: AXUIElement tree (buttons, links, fields)
    ├── Actions: pyautogui (mouse, keyboard, scroll)
    └── Security: PII regex filter (API keys, cards, passwords, emails)

The loop:

1. Look — vision_read_window("Safari") → text + coordinates

2. Decide — agent plans next action

3. Act — action_click_text("Sign In") → clicks center of text

4. Verify — vision_read_window again → confirm it worked

5. Repeat

Features

Native macOS OCR

Uses Apple Vision Framework — no external API, no GPU needed, works offline. Supports Russian and English.

Per-Window Capture

Captures specific windows via CGWindowList without switching apps or stealing focus.

Smart Coordinates

OCR returns absolute screen coordinates with Retina scaling handled automatically.

OCR Cache

Repeated reads of unchanged windows are 7x faster. Cache auto-invalidates after any action.

PII Protection

Sensitive data is automatically masked in OCR output:

• API keys: sk-1***ef

• Credit cards: 4111***1111

• Emails: user***com

• Passwords in forms

• Crypto keys

Accessibility API

Detect interactive UI elements — buttons, checkboxes, links, text fields — with exact clickable coordinates.

Built-in Agent Instructions

The MCP server includes a system prompt that teaches agents how to use all 27 tools, macOS keyboard shortcuts, and the look-decide-act-verify loop.

Security

• All data stays on your Mac — no telemetry, no analytics, no external calls

• PII auto-masking — API keys, credit cards, emails, passwords masked before reaching the agent

• URL validation — only http:// and https:// allowed, file:// blocked

• Input sanitization — app names validated to prevent injection

• Browser allowlist — only known browsers accepted (Chrome, Safari, Firefox, Arc, etc.)

• Failsafe — pyautogui failsafe enabled by default (move mouse to corner to abort)

Examples

import asyncio
from puppet_ai.core.capture import ScreenCapture
from puppet_ai.core.actions import DesktopActions
from puppet_ai.server.mcp import VisionPipeContext, create_all_tools

async def main():
    ctx = VisionPipeContext(
        capture=ScreenCapture(),
        actions=DesktopActions(failsafe=True),
    )
    tools = create_all_tools(ctx)

    # See what's on screen
    windows = await tools["vision_list_windows"]()
    for w in windows:
        print(f"{w['app']:20s} — {w['title'][:50]}")

    # Read a window
    page = await tools["vision_read_window"](app="Safari")
    print(page["text"][:500])

    # Click text on screen
    await tools["action_click_text"](text="Sign In", app="Safari")

    # Open a URL
    await tools["action_open_url"](url="https://example.com", browser="Safari")

asyncio.run(main())

More examples in examples/.

Configuration

# puppet-ai.yaml
ocr:
  languages: ["en", "ru"]
  mode: accurate  # or "fast"

pii:
  enabled: true
  categories: [api_keys, credit_cards, crypto_keys, emails, passwords]

capture:
  max_width: 800
  format: jpeg
  quality: 75

Presets:

puppet-ai serve --preset fast      # speed over accuracy
puppet-ai serve --preset balanced  # default
puppet-ai serve --preset quality   # max accuracy

Known Limitations

• Safari — type_text and Cmd+V don't work reliably in web input fields. Safari restricts programmatic input and Accessibility API access for web content. Use Chrome for full browser automation.

• Electron apps — some Electron apps (e.g. Dolphin Anty) block clipboard paste in input fields.

• Safari Accessibility — returns 0 UI elements for web page content. Native macOS apps and Chrome work fine.

For full browser automation, Chrome is recommended over Safari.

Requirements

• macOS 13+ (Ventura or later)

• Python 3.11+

• Accessibility permissions enabled

Author

Daniel Starkov

• Twitter: @retardTransoff

• LinkedIn: Daniel Starkov

License

MIT