Microsoft Teams Bridge for Cartesia Line Agents (Python)

Welcome! cartesia-msteams-bridge (PyPI: cartesia-msteams-bridge) puts a Cartesia Line voice agent on a real Microsoft Teams call.

It is a small Python service (and importable asyncio library, built on aiohttp) that sits between two WebSockets:

flowchart LR
    Teams["Microsoft Teams<br/>call"]
    StandIn["StandIn media bridge<br/>(hosted service)"]
    Bridge["this bridge<br/>(yours)"]
    Line["Cartesia Line agent<br/>your agent code on<br/>Cartesia's platform"]
    Teams <--> StandIn
    StandIn -- "HMAC WS, base64 PCM 16k" --> Bridge
    Bridge -- "WS, base64 pcm_16000" --> Line

The hosted StandIn media bridge (standin.komaa.com) joins the Teams call and dials into your bridge, one authenticated WebSocket per call. The bridge opens one Line agent stream per call (via Cartesia’s WebSocket API, the integration Cartesia provides for bringing your own telephony) and relays between them. Both wires carry base64 PCM 16 kHz mono, so the hot path relays the payload string verbatim: no decode, no re-encode, no resampling, no transcoding.

Where the brain lives

A Line agent is your code deployed on Cartesia’s platform: the LLM, the tools, the conversation logic, the hangup decision. That makes this bridge deliberately a transport - unlike the sibling bridges there is no in-process tool registry and no vision hook, because the Line wire has no client-side tool channel. Your agent still gets the full Teams context (caller identity, participants, active speaker, DTMF, recording state) through the wire’s metadata and custom events - see Your Line Agent.

What it does

Realtime voice - the caller talks to your Line agent and hears it reply. Turn-taking and interruption are the Line platform’s own; the bridge maps Line’s clear onto a Teams-side playback flush.
Per-call personalization - caller name, tenant and direction ride the start metadata for your agent code; CARTESIA_INTRODUCTION gives a deterministic opening line (a natural spoken AI disclosure), CARTESIA_VOICE_ID overrides the voice, CARTESIA_SYSTEM_PROMPT optionally overrides the prompt (with caller context appended).
Native DTMF - keypad digits are forwarded as Line dtmf events, not prompt hacks.
Live context events - participants, active-speaker changes and recording state reach your agent code as custom events, with an initial snapshot at stream start.
Per-call access tokens - the long-lived API key never rides an agent socket; each call mints a short-lived, agent-scoped token.
Two call governors - StandIn-side tier cutoffs get a spoken goodbye; a bridge-side MAX_CALL_MINUTES hard cap protects your budget. With Sonic TTS configured the goodbye is the exact text, deterministic and backstopped.
Hardened transport - replay-proof HMAC handshake, connection and payload caps, duplicate-call rejection, dead-peer detection, ack-gated audio, graceful drain, and a *.cartesia.ai host allowlist so your API key can never be sent elsewhere.

Use the sidebar to navigate. Start with Getting Started, then read Your Line Agent for what your agent code receives. There is a runnable example project in the repository - walked through in Run the Example.