pipecat

Pipecat MCP Server exposes voice-related and screen capture capabilities to MCP-compatible clients. It uses local, on-device models by default (Faster Whisper for speech-to-text and Kokoro for text-to-speech) so no API keys are required, and the initial startup may take a moment as Whisper models are downloaded. The server is designed to work with standard MCP clients, allowing you to control the agent and access screen analysis tools through MCP tool calls while audio transport is handled by separate means such as Pipecat Playground, Daily, or phone providers. To start a session, run the Pipecat MCP Server, then connect an MCP client to http://localhost:9090/mcp and use an audio transport to communicate with the agent.

From there, you can use MCP clients like Claude Code, Cursor, or Codex to issue commands to the agent. If you want a smoother voice experience, you can install the recommended Talk Skill, auto-approve permissions, and enable an audio transport (Playground, Daily, or phone-based) to hear and talk with the agent. The server also supports screen capture and analysis through the Screen Capture & Analysis features, enabling the agent to help debug issues or provide feedback on UI elements by analyzing the shared screen.

Prerequisites:

• Python 3.10 or later
• uv package manager

Install the MCP server tool via uv:

uv tool install pipecat-ai-mcp-server

This installs the pipecat-mcp-server CLI that you will run to start the MCP server.

If you want to modify or customize the server locally, clone the repository and install your local version:

git clone https://github.com/pipecat-ai/pipecat-mcp-server.git

uv tool install -e /path/to/repo/pipecat-mcp-server

Running the server after installation:

pipecat-mcp-server

The server will be available at http://localhost:9090/mcp.

Environment variables and configuration tips:

• DAILY integration (optional): install with Daily support via uv tool features and set DAILY_API_KEY and DAILY_ROOM_URL when using the Daily prebuilt transport. Example: export DAILY_API_KEY=your-daily-api-key export DAILY_ROOM_URL=your-daily-room pipecat-mcp-server -d
• Screen capture: The server relies on platform-compatible screen capture tools (ScreenCaptureKit on macOS, Xlib on Linux). Ensure you have permissions to capture screens where you intend to analyze UI.
• Auto-approving permissions: For hands-free operation, configure tool permissions to avoid prompts. This may be required for certain MCP clients.
• Talk skill (recommended): Installing the talk skill improves voice interaction by asking for verbal confirmation before making file changes.
• Audio transport: The MCP server provides the voice tools, but audio input/output is handled by separate transport layers (Pipecat Playground, Daily, or phone providers). Ensure your transport is properly connected to hear and speak with the agent.