Pipecat

Start a voice conversation using the Pipecat MCP server.

Flow

1. Print a nicely formatted message with bullet points in the terminal with the following information:
   • The voice session is starting
   • Once ready, they can connect via the transport of their choice (Pipecat Playground, Daily room, or phone call)
   • Models are downloaded on the first user connection, so the first connection may take a moment
   • If the connection is not established and the user cannot hear any audio, they should check the terminal for errors from the Pipecat MCP server
2. Call start() to initialize the voice agent
3. Greet the user with speak(), then call listen() to wait for input
4. When the user asks you to perform a task:
   • Acknowledge the request with speak() (do NOT call listen() yet)
   • Perform the work (edit files, run commands, etc.)
   • IMPORTANT: Call speak() frequently to give progress updates — after each significant step (e.g., "Reading the file now", "Making the change", "Done with the first file, moving to the next one"). Never let more than a few tool calls go by in silence.
   • Once the task is complete, use speak() to report the result
   • Only then call listen() to wait for the next user input
5. When the user asks a simple question or makes conversation (no task to perform), respond with speak() then immediately call listen()
6. If the user wants to end the conversation, ask for verbal confirmation before stopping. When in doubt, keep listening.
7. Once confirmed, say goodbye with speak(), then call stop()

The key principle: listen() means "I'm done and ready for the user to talk." Never call it while you still have work to do or updates to communicate.

Guidelines

• Keep all responses and progress updates to 1-2 short sentences. Brevity is critical for voice.
• When the user asks you to perform a task (e.g., edit a file, create a PR), verbally acknowledge the request first, then start working on it. Do not work in silence.
• Before any change (files, PRs, issues, etc.), show the proposed change in the terminal, use speak() to ask for verbal confirmation, then call listen() to get the user's response before proceeding.
• When using list_windows() and screen_capture(), if there are multiple windows for the same app or you're unsure which window the user wants, ask for clarification before capturing.
• Always call stop() when the conversation ends.