AI-powered design tool access for self-hosted Penpot via Model Context Protocol.

An MCP server that gives AI agents (like Claude Code, Cursor, or any MCP-compatible client) full programmatic access to your self-hosted Penpot instance. AI can read, create, modify, and export design elements — from rectangles and text to full UI components — all through natural language.

Think of it as the bridge between your AI assistant and your design tool.

graph TB
    AI["AI Agent\n(Claude Code · Cursor · Gemini CLI)"]

    subgraph SERVERS["MCP Layer"]
        MCP["penpot-mcp — Python\n68 tools · :8787\nDB reads + API writes + Plugin"]
        OMCP["Penpot MCP — Official\n~20 tools · penpot/penpot monorepo\nPlugin API only · TypeScript"]
    end

    subgraph PENPOT["Penpot Stack (Docker)"]
        PG["PostgreSQL\n:5432"]
        BE["Backend\n:6060"]
        FE["Frontend\n:9001"]
        EX["Exporter\n:6061"]
    end

    subgraph BRIDGE["Browser Plugin Bridge"]
        WS["WebSocket Server\n:4402"]
        UI["ui.html\niframe · full browser API"]
        PJ["plugin.js\nworker sandbox · penpot.*"]
    end

    AI -->|"Streamable HTTP :8787"| MCP
    AI -->|"Streamable HTTP"| OMCP

    MCP -->|"asyncpg · direct SQL"| PG
    MCP -->|"httpx · RPC API"| BE
    MCP -->|"PNG / SVG export"| EX
    MCP <-->|"WebSocket"| WS

    OMCP <-->|"WebSocket :4402"| WS

    WS <-->|"ws://localhost:4402"| UI
    UI <-->|"postMessage"| PJ
    PJ -->|"penpot.* API"| FE
    FE -.->|"proxy"| BE
    BE --> EX

    style AI fill:#7c3aed,color:#fff
    style MCP fill:#2563eb,color:#fff
    style OMCP fill:#0f766e,color:#fff
    style PG fill:#16a34a,color:#fff
    style BE fill:#ea580c,color:#fff
    style FE fill:#ea580c,color:#fff
    style EX fill:#ea580c,color:#fff
    style WS fill:#0891b2,color:#fff
    style UI fill:#0891b2,color:#fff
    style PJ fill:#0891b2,color:#fff

Tri-layer access strategy:

• Reads go directly to PostgreSQL via asyncpg — fast and reliable, bypasses API overhead
• Writes go through Penpot's RPC API via httpx — ensures proper change tracking and undo history
• Exports use Penpot's built-in exporter (headless Chromium) for pixel-perfect SVG/PNG output
• Live canvas goes through the Browser Plugin bridge (port 4402) — shared architecture with the official Penpot MCP, enabling both servers to coexist and complement each other in the same AI workflow

1. Self-hosted Penpot running via Docker Compose (official guide)
2. Docker and Docker Compose v2 installed
3. Access tokens enabled in your Penpot instance (see Enable Access Tokens)

git clone https://github.com/ancrz/penpot-mcp-server.git
cd penpot-mcp-server
chmod +x setup.sh
./setup.sh

The script will guide you through configuration, build the Docker image, and start the server.

git clone https://github.com/ancrz/penpot-mcp-server.git
cd penpot-mcp-server

cp .env.example .env

Edit .env with your Penpot details:

# Your Penpot access token (see "Enable Access Tokens" below)
PENPOT_ACCESS_TOKEN=your-token-here

# Your Penpot database password (from your Penpot docker-compose.yml)
PENPOT_DB_PASS=your-db-password

# Public URL where you access Penpot in the browser
PENPOT_PUBLIC_URL=http://localhost:9001

Add the penpot-mcp service definition to your existing Penpot docker-compose.yml. See docker-compose.penpot.yml for the complete service definition to copy.

docker compose up -d --build penpot-mcp

# Quick health check
curl -s http://localhost:8787/
# → {"service": "Penpot MCP", "status": "ok", "version": "0.1.0"}

# Full MCP protocol initialization
curl -s http://localhost:8787/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

You should see a JSON response with the server capabilities.

Penpot MCP uses network transport (streamable HTTP) — the server runs as a Docker container and clients connect via HTTP. This means:

• The server is always running independently (via Docker Compose)
• The client only needs the URL to connect — no process spawning
• env in the client's JSON config is irrelevant — credentials live in the server's own .env file (configured during setup)
• Any client on the same machine (or network) can connect to http://localhost:8787/mcp

Key difference from stdio servers: With stdio servers (like Skill Swarm), the client launches the process and injects env vars. With network servers like Penpot MCP, the server manages its own credentials. The env block in your client's MCP config has no effect.

Claude Code uses "type": "http" for streamable HTTP connections.

Global (~/.claude.json):

{
  "mcpServers": {
    "penpot": {
      "type": "http",
      "url": "http://localhost:8787/mcp"
    }
  }
}

Project-level (.mcp.json in your project root):

{
  "mcpServers": {
    "penpot": {
      "type": "http",
      "url": "http://localhost:8787/mcp"
    }
  }
}

Restart Claude Code. You should see 68 tools from the penpot server listed when you run /mcp.

Note: Use "type": "http", not "streamable-http". Claude Code maps http to the streamable HTTP transport internally. Using streamable-http will cause a schema validation error.

Gemini CLI uses httpUrl (not url) for streamable HTTP connections. Transport is inferred from the field name.

Config file: ~/.gemini/settings.json

{
  "mcpServers": {
    "penpot": {
      "httpUrl": "http://localhost:8787/mcp"
    }
  }
}

Note: Gemini CLI distinguishes between url (SSE transport) and httpUrl (streamable HTTP transport). Penpot MCP uses streamable HTTP, so use httpUrl. No type field needed.

Antigravity uses serverUrl for HTTP-based MCP servers.

Config file: ~/.gemini/antigravity/mcp_config.json

{
  "mcpServers": {
    "penpot": {
      "serverUrl": "http://localhost:8787/mcp"
    }
  }
}

Note: Antigravity uses serverUrl (not url or httpUrl). If Antigravity runs inside Docker, make sure it can reach localhost:8787 on the host — you may need host.docker.internal:8787 instead of localhost:8787 depending on your Docker network setup.

Once connected, you can ask your AI agent things like:

• "List my Penpot projects"
• "Create a login form with email/password fields and a submit button"
• "Export the Login Card frame as SVG"
• "What colors are defined in the design system?"
• "Add a comment at position (100, 200) saying 'Review this layout'"

The Penpot MCP Plugin bridges the AI agent with the live Penpot canvas, enabling real-time context awareness:

• Live selection: AI can query which shapes you currently have selected
• Script execution: AI can run JavaScript directly via the Penpot Plugin API

These features require the browser plugin to be connected. The 66 headless tools work without it.

1. Make sure the MCP server is running: docker compose up -d penpot-mcp
2. Open Penpot in your browser
3. Press Ctrl+Alt+P (or Main Menu -> Plugin Manager)
4. Paste the URL in the input field: http://localhost:8787/plugin/manifest.json
5. Click Install → Allow on the permissions dialog
6. Click Open to launch the plugin panel

The plugin panel appears on the right. When the status indicator turns green, the AI agent has live access to the canvas.

The Penpot backend must have enable-plugins-runtime in PENPOT_FLAGS:

PENPOT_FLAGS=enable-login-with-password enable-registration enable-access-tokens enable-plugins-runtime

Restart required: After adding enable-plugins-runtime, restart both penpot-backend and penpot-frontend:

docker compose restart penpot-backend penpot-frontend

Chrome may show a permission popup: "Allow [localhost:9001] to access your local network?"

1. Click Allow when the popup appears
2. The plugin will connect automatically

If no popup appears and the plugin stays disconnected, check chrome://flags/#private-network-access-respect-preflight-results -- disable it for local development.

1. Click the Shield icon (lion) in the address bar
2. Disable the Shield for localhost:9001 (or set to "No Blocking")
3. Reload the Penpot tab and reconnect the plugin

The server provides 68 tools across 11 categories. See TOOLS.md for the complete reference with all parameters.

All settings are via environment variables. See .env.example for a template.

Penpot requires a feature flag to enable API access tokens.

Add enable-access-tokens to your PENPOT_FLAGS:

PENPOT_FLAGS=enable-login-with-password enable-registration enable-access-tokens

docker compose restart penpot-backend penpot-frontend

1. Open Penpot in your browser
2. Click your avatar (bottom-left) → Access Tokens
3. Click "Generate new token"
4. Give it a name (e.g., "MCP Server")
5. Copy the token and paste it into your .env as PENPOT_ACCESS_TOKEN

The MCP server runs as a Docker container alongside your existing Penpot stack. You need to add it to your Penpot docker-compose.yml.

See docker-compose.penpot.yml for the exact service definition to add. The key points:

• It connects to the penpot Docker network (same as other Penpot services)
• It depends on penpot-postgres (with health check) and penpot-backend
• It exposes port 8787 on localhost only (127.0.0.1:8787:8787)
• Environment variables reference Docker internal hostnames

# Install uv if needed
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies
uv sync

# Run the server (needs .env configured for local access)
uv run penpot-mcp

For local development, point PENPOT_DB_HOST and PENPOT_DB_PORT to your host-mapped PostgreSQL port, and PENPOT_BASE_URL to http://localhost:9001.

uv sync --group dev
uv run pytest tests/ -v

penpot-mcp-server/
├── src/penpot_mcp/
│   ├── server.py            # FastMCP entry point, 68 tool registrations, plugin routes
│   ├── config.py            # Pydantic Settings configuration
│   ├── gateway.py           # Hybrid context gateway (DB + Plugin awareness)
│   ├── ws_controller.py     # WebSocket server for browser plugin bridge (:4402)
│   ├── plugin/
│   │   ├── manifest.json    # Penpot plugin manifest (served at /plugin/manifest.json)
│   │   ├── plugin.js        # Plugin worker — penpot.* API only (no WebSocket in sandbox)
│   │   └── ui.html          # Plugin iframe — WebSocket lives here, relays to plugin.js
│   ├── services/
│   │   ├── db.py            # asyncpg connection pool
│   │   ├── api.py           # httpx RPC API client
│   │   ├── changes.py       # Penpot change operations builder
│   │   └── transit.py       # Transit+JSON decoder
│   ├── tools/
│   │   ├── projects.py      # Team & project queries
│   │   ├── files.py         # File CRUD operations
│   │   ├── shapes.py        # Shape reading & search
│   │   ├── create.py        # Shape creation
│   │   ├── modify.py        # Shape modification
│   │   ├── text.py          # Text operations
│   │   ├── export.py        # PNG/SVG export
│   │   ├── components.py    # Components & design tokens
│   │   ├── comments.py      # Comments & collaboration
│   │   ├── media.py         # Media assets & fonts
│   │   ├── database.py      # Raw SQL queries
│   │   └── advanced.py      # File raw data & revision comparison
│   └── transformers/
│       ├── css.py           # Shape → CSS conversion
│       ├── svg.py           # Shape → SVG conversion
│       └── layout.py        # Layout → CSS flexbox/grid
├── tests/
│   ├── conftest.py
│   ├── test_projects.py
│   ├── test_files.py
│   ├── test_shapes.py
│   └── test_e2e_login_form.py
├── pyproject.toml
├── Dockerfile
├── .env.example
├── setup.sh
├── docker-compose.penpot.yml
├── TOOLS.md
└── LICENSE

This project is licensed under the Apache License 2.0.

• Penpot — The open-source design platform
• Model Context Protocol — The protocol standard
• FastMCP — Python MCP SDK