- Your Pipecat bot on an Unpod call. Unpod carries the phone call and streams its audio to your bot; your pipeline does STT, the LLM and TTS. Jump to the walkthrough.
- The Unpod brain inside your Pipecat pipeline, below.
FrameProcessor. Audio that Pipecat captures
flows through your DialogMachine, and the reply streams back out the same pipeline.
Pipecat is a runtime, not a SIP provider. There’s no Dashboard/API provider
config - you wire it in code. The full, working integration lives in the embedding guide.
Pipecat embedding guide
The drop-in
FrameProcessor, install steps, and a complete pipeline example.How it fits
🎙️Mic / SIP audioAny Pipecat transport
① frames
⚙️Pipecat pipelineSTT · VAD · TTS stages
② user turn③ reply tokens
🧠Unpod FrameProcessorDialogMachine
④ TTS stream
🔊Audio outLow-latency reply
- Bring audio in - any Pipecat transport (WebRTC, Daily, telephony) feeds frames.
- Insert the brain - add the Unpod
FrameProcessorso each user turn drives yourDialogMachine. - Stream the reply - tokens stream back through Pipecat’s TTS stage for low latency.
Your Pipecat bot on an Unpod call
Unpod answers the call and asks your answer URL what to do. You reply with<Stream>, Unpod
opens a WebSocket to your bot, and UnpodFrameSerializer turns its messages into Pipecat
frames and back. The rest of your bot is plain Pipecat. The protocol underneath is on
Audio Streaming.
Preview. Your org needs the Call API and audio streaming switched on. Ask the Unpod team.
1
Credentials and install
Export The bot below uses Deepgram, Anthropic and Cartesia; export
UNPOD_BASE_URL, UNPOD_PLATFORM_TOKEN and UNPOD_ORG_HANDLE as shown in
Credentials. Python 3.12+.DEEPGRAM_API_KEY,
ANTHROPIC_API_KEY and CARTESIA_API_KEY, or swap in any Pipecat services.2
Get a number and your webhook secret
Attach a number with
attach_type="human_call" so your app answers it. Then get your
org’s webhook secret: create it once
with webhook_secret.rotate(), or read the current one with webhook_secret.reveal()
(org admin). Store it as UNPOD_WEBHOOK_SECRET. Unpod signs the stream’s WebSocket
upgrade with it, so your bot only accepts streams from Unpod.3
Run the bot
One app, two routes: Run it with
/answer replies with <Stream>, and /ws is the stream.
Save as app.py:app.py
python app.py.4
Expose it
Unpod only calls public https URLs, and streams to a public Set
ws:// or wss:// host
(use wss:// in production). In a second terminal:PUBLIC_HOST to the ngrok host (e.g. abc123.ngrok.app) and restart app.py.5
Point your number at it, or call out
Inbound, once per number:Outbound, per call:
6
Talk to it
Call your number, or pick up the outbound call. The bot greets you; talk and it answers.
When the caller hangs up, Unpod closes the socket and the pipeline stops. When the bot
ends (an
EndFrame), the serializer hangs up the call.POST /start route protected by an
X-Api-Key header, is in the SDK repo:
examples/pipecat_streaming.
What UnpodFrameSerializer does
- It reads the format from
start: μ-law or A-law at 8000 Hz, or L16 at 8000, 16000 or 24000 Hz. Run the pipeline at 8000 Hz for μ-law and A-law calls to skip resampling. - It feeds only the caller’s (
inbound) audio to the pipeline, so the bot never hears itself even withaudioTrack="both". - Hanging up needs
client=(anAsyncClient) or your ownhangup=coroutine. A call that already ended (404, 409, 410) is ignored. Turn it off withparams=UnpodFrameSerializer.InputParams(auto_hang_up=False).
When to pick Pipecat
Bring your agent
The adapter pattern any runtime uses.
All platforms
Back to the integrations overview.