DRU-660 -- The calls server: Pipecat carries the call

[czpython]

What changed

The calls server carries a phone call. It takes Twilio's audio stream, asks Druks to pick up the call, holds the call with the speech-to-speech model, runs the Bot's tools through /mcp, and posts each line to Druks. It is its own program on Pipecat, with its own image and its own release cycle.

• The project. druks-calls/ holds the package calls, with its own pyproject.toml and uv.lock. It pins pipecat-ai 1.12.0 with the Google, MCP, OpenAI, and WebSocket extras. The backend does not depend on it, and it imports nothing from Druks.
• The server. app.py serves GET /healthz and a WebSocket at /_calls/{token}. At the handshake it posts pickup to {DRUKS_URL}/_external/calls/events/ with the token and the stream's X-Twilio-Signature. It closes the stream when Druks refuses, and it refuses a stream beyond MAX_CALLS live calls.
• A call. Call.run builds one pipeline for each call: the Twilio transport, the realtime service of the model's vendor (openai/… or google/…), and one MCP client with the pickup's key and conversation header. The model gets the pickup's prompt, then the facts, then the tools. A model that cannot connect ends the call.
• The lines. The context aggregators give the caller's and the assistant's lines. Pipecat adds the caller's words after the reply has started, so each assistant turn numbers the caller's line first and the reply second. The server posts each line under its number, and Druks orders a call by those numbers.
• The end. end_call is the model's tool to end the call. A call also ends after no_speech_seconds of silence, at max_call_seconds, or when the caller hangs up. The server then posts ended.
• Names on the Druks side. The server's webhook is CallsEvents at /_external/calls/events/, and a call's stream URL is wss://<webhook host>/_calls/<call token>. Twilio.link takes the address for a number's calls as calls_url.
• The Voice card. It refuses a model whose vendor is not openai or google, because the calls server runs only those two. It has a new optional field, the transcription model, which an OpenAI model uses to write the caller's words as text.
• Its own release cycle. deploy/calls/Dockerfile builds the image from druks-calls/, with goss.yaml beside it. build-calls-image.yml publishes ghcr.io/czpython/druks-calls only when the server changes. on-pull-request-calls.yml runs its tests, builds the image, and runs the goss checks, also only when it changes. The backend PR workflow lints druks-calls/.
• Not in the default stack. Most installs do not need calls, so deploy/compose.yaml has no calls service and druks.toml has no switch for it. docs/deployment.md gives the service to add to compose.override.yaml. A Druks upgrade restarts it only when a new calls image exists.
• Caddy. deploy/caddy/Caddyfile passes /_calls/* to the calls server, on the webhook host and on the :8000 listener.
• Docs. docs/deployment.md covers the calls server and how to add it, docs/full-local.md the tunnel for calls, docs/chat.md when a call ends, and AGENTS.md the project.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
druks	🟢 Ready	View Preview	Oct 4, 2026, 2:19 PM

💡 Tip: Enable Automations to automatically generate PRs for you.