pipecat/examples/multi-task

Pipecat Multi-Task Examples

This directory contains example bots that use the multi-task framework in pipecat.tasks, pipecat.pipeline.runner (with spawn()), and the TaskBus. Each example shows a different cooperation pattern between tasks: hand-off, parallel fan-out, remote workers, etc.

Setup

From the repo root:

uv sync --all-extras
source .venv/bin/activate
cd examples/multi-task

Copy the env template and fill in your API keys:

cp env.example .env

Environment variables

Variable	Required by
OPENAI_API_KEY	LLM tasks
DEEPGRAM_API_KEY	STT
CARTESIA_API_KEY	TTS
DAILY_API_KEY	Optional: only with --transport daily

Additional, example-specific variables are listed below.

Table of contents

Local (single process)

• Handoff between LLM tasks
• Parallel debate
• Voice code assistant with Claude Agent SDK

Distributed (multi-process)

• Handoff via Redis
• Handoff via PGMQ (Postgres)
• LLM task via WebSocket proxy

Local

Examples where all tasks run in the same process on an AsyncQueueBus.

Handoff between LLM tasks

Two LLM tasks (greeter + support) that transfer control to each other during a voice conversation. A main task owns the transport pipeline and bridges frames to the bus.

Running

uv run local-handoff/local-handoff-two-agents.py

Open http://localhost:7860/client in your browser to talk to your bot.

To use Daily transport:

uv run local-handoff/local-handoff-two-agents.py --transport daily

Overview

• local-handoff-two-agents.py — Two LLM tasks (greeter + support) that hand off via BaseTask.handoff_to(...). The main task owns STT, TTS, transport, and a BusBridgeProcessor.
• local-handoff-two-agents-tts.py — Same shape, but each child task ships with its own CartesiaTTSService in a custom pipeline. The main task has no TTS — audio comes from whichever child is active over the bus.

Parallel debate

Parallel fan-out using task.job_group(...). A voice bot takes a topic from the user, kicks off three worker tasks in parallel (advocate, critic, analyst), waits for all three to respond, and synthesizes a balanced answer. Each worker keeps its own LLM context across rounds.

Running

uv run parallel-debate/parallel-debate.py

Open http://localhost:7860/client in your browser to talk to your bot.

To use Daily transport:

uv run parallel-debate/parallel-debate.py --transport daily

Architecture

Main task (transport + LLM + `debate` tool)
  └── job_group(advocate, critic, analyst)
        └── DebateWorker (LLMContextTask, one per role)

• Main task: transport (STT, TTS) + LLM moderator with a debate direct function that fans out via task.job_group(...).
• Debate workers: LLMContextTasks spawned on the runner. Each keeps its own LLMContext across rounds and ships its completed turn back as a job response via the assistant-aggregator's on_assistant_turn_stopped event.

Voice code assistant

Talk to your codebase hands-free. Ask questions about code, project structure, or file contents and get spoken answers based on actual files. The Claude Agent SDK worker navigates the filesystem using Read, Bash, Glob, and Grep tools.

Additional environment variables

Variable	Required by
ANTHROPIC_API_KEY	Code worker (Claude Agent SDK)
PROJECT_PATH	Optional, defaults to cwd

Running

# Default: explores the current directory
uv run code-assistant/code-assistant.py

# Specify a project path
PROJECT_PATH=/path/to/your/project uv run code-assistant/code-assistant.py

Open http://localhost:7860/client in your browser to talk to your bot.

To use Daily transport:

uv run code-assistant/code-assistant.py --transport daily

Example questions

• "What does the main module do?"
• "Find all TODO comments in the project"
• "How is error handling implemented?"
• "What dependencies does this project use?"
• "Explain the test structure"

Architecture

Main task (transport + LLM + `ask_code` tool)
  └── job → CodeWorker (Claude Agent SDK)

• code-assistant.py — Main task: STT, LLM (with system prompt + ask_code direct function), TTS, and transport. The ask_code tool dispatches a job to the worker via task.job("code_worker", payload=...).
• code_worker.py — CodeWorker: a bus-only BaseTask spawned on the runner. It accepts @job-style requests through the bus and runs them sequentially through a persistent Claude SDK session so follow-up questions share context.

Distributed

Examples where tasks run across separate processes or machines.

Handoff via Redis

Same two-task handoff as the local example, but each task runs as a separate process connected via Redis pub/sub. Requires pip install pipecat-ai[redis].

Quick start (single machine, local Redis)

Terminal 1: start Redis

docker run --rm -p 6379:6379 redis:7

Terminal 2: start the greeter worker

uv run distributed-handoff/redis-handoff/llm.py greeter

Terminal 3: start the support worker

uv run distributed-handoff/redis-handoff/llm.py support

Terminal 4: start the main transport task

uv run distributed-handoff/redis-handoff/main.py

All processes connect to redis://localhost:6379 by default.

Running across machines

Point each process at the same Redis instance:

Machine A

uv run distributed-handoff/redis-handoff/main.py --redis-url redis://your-redis-host:6379

Machine B

uv run distributed-handoff/redis-handoff/llm.py greeter --redis-url redis://your-redis-host:6379

Machine C

uv run distributed-handoff/redis-handoff/llm.py support --redis-url redis://your-redis-host:6379

Architecture

Machine A                    Redis                    Machine B
+------------+          +-------------+          +-------------+
| main.py    |  <---->  | pub/sub     |  <---->  |    llm.py   |
| (transport,|          | channel:    |          |  (greeter)  |
|  STT, TTS) |          | pipecat:acme|          +-------------+
+------------+          +-------------+          +-------------+
                               ^                 |    llm.py   |
                               +-------------->  |  (support)  |
                                                 +-------------+

• main.py — Transport task: Daily/WebRTC, Deepgram STT, Cartesia TTS, and a BusBridgeProcessor over a RedisBus.
• llm.py — LLM worker: runs either greeter or support with OpenAI behind a bridged LLMTask.

Handoff via PGMQ (Postgres)

Same shape as the Redis handoff, but the bus is backed by PGMQ on a shared Postgres database (e.g. Supabase). Requires pip install pipecat-ai[pgmq].

Additional environment variables

Variable	Required by
DATABASE_URL	PostgreSQL DSN (e.g. Supabase pooled connection string)
PGMQ_CHANNEL	Optional, channel prefix for queue names. Defaults to pipecat_acme

Quick start

Terminal 1: start the greeter worker

uv run distributed-handoff/pgmq-handoff/llm.py greeter --database-url $DATABASE_URL

Terminal 2: start the support worker

uv run distributed-handoff/pgmq-handoff/llm.py support --database-url $DATABASE_URL

Terminal 3: start the main transport task

uv run distributed-handoff/pgmq-handoff/main.py --database-url $DATABASE_URL

You can also set DATABASE_URL in .env and omit the --database-url flag.

Architecture

Same as the Redis handoff above; the RedisBus is replaced by a PgmqBus, and the "pub/sub channel" is a set of PGMQ queues on the shared Postgres instance.

LLM task via WebSocket proxy

Runs an LLM task on a remote server, connected to the main transport task via a WebSocket proxy. No shared bus required — the proxy tasks forward bus messages point-to-point over the WebSocket.

Quick start (single machine)

Terminal 1: start the remote assistant server

uv run remote-proxy-assistant/assistant.py

Terminal 2: start the main transport task

uv run remote-proxy-assistant/main.py --remote-url ws://localhost:8765/ws

Open http://localhost:7860/client in your browser to talk to the bot.

Running across machines

Server machine: start the assistant

uv run remote-proxy-assistant/assistant.py --host 0.0.0.0 --port 8765

Client machine: point at the server

uv run remote-proxy-assistant/main.py --remote-url ws://server-host:8765/ws

Architecture

    +-------------+    +-------------+           +-------------+     +-----------------+
    |             |    |             |           |             |     |                 |
    |  Main task  |    | Proxy task  |  <~~~~~>  | Proxy task  |     | Assistant task  |
    |             |    |  (client)   |           |  (server)   |     |                 |
    +-------------+    +-------------+           +-------------+     +-----------------+
        messages           messages                  messages             messages
            │                 │                         │                    │
  ══════════╧═════════════════╧════════         ════════╧════════════════════╧═══════════
                Task Bus                                       Task Bus
  ═════════════════════════════════════         ═════════════════════════════════════════

• main.py — Transport task with STT, TTS, and a BusBridge. Spawns a WebSocketProxyClientTask that connects to the remote server and forwards BusFrameMessages.
• assistant.py — FastAPI server. Each WebSocket connection spawns a WebSocketProxyServerTask plus a bridged AcmeAssistant LLM task on a per-session PipelineRunner.

Security

The proxy tasks filter messages by task name:

• Only messages targeted at the remote task cross the WebSocket
• Only messages targeted at the local task are accepted from the WebSocket
• Broadcast messages never cross the WebSocket

Pass HTTP headers for authentication:

proxy = WebSocketProxyClientTask(
    "proxy",
    url="wss://server-host:8765/ws",
    remote_task_name="assistant",
    local_task_name="acme",
    headers={"Authorization": "Bearer <token>"},
)