Getting Started

By the end of this page a Deepgram Voice Agent answers a Microsoft Teams call. You need Python >= 3.10, a Deepgram API key, and a StandIn identity (the sandbox is enough).

There is no agent to set up in a dashboard: the bridge configures each Voice Agent session itself (STT model, LLM, voice, prompt, greeting, tools) from your environment variables.

1. Install and run the bridge

pip install deepgram-msteams-bridge

As a CLI:

DEEPGRAM_API_KEY=dg_... \
WORKER_SHARED_SECRET=... \
  deepgram-msteams-bridge

A .env file in the working directory is loaded automatically (existing environment wins). Or embedded in your own asyncio app:

import asyncio
from deepgram_msteams_bridge import load_config, start_server

async def main():
    await start_server(load_config())  # same env variables as the CLI
    await asyncio.Event().wait()

asyncio.run(main())

Every option is an environment variable; the package ships a fully commented .env.example, and the Configuration Reference documents each one. The bridge listens on 0.0.0.0:8080 by default and exposes GET /healthz for liveness checks.

WORKER_SHARED_SECRET comes from StandIn in the next step.

2. Shape the agent (optional, recommended)

The defaults work out of the box (nova-3 STT, open_ai/gpt-4o-mini thinking, aura-2-thalia-en voice); these variables define the personality:

DEEPGRAM_PROMPT="You are Komaa's friendly receptionist. Keep replies short; you are speaking aloud on a phone call."
DEEPGRAM_GREETING="Hello! You've reached Komaa. How can I help?"
DEEPGRAM_SPEAK_MODEL=aura-2-thalia-en
DEEPGRAM_THINK_MODEL=gpt-4o-mini

The bridge appends per-call caller context (name, tenant, direction) to your prompt automatically, and registers the built-in tools (end_call, look, show_image, express) on every session. Third-party LLMs (google, groq, aws_bedrock) additionally need DEEPGRAM_THINK_ENDPOINT_URL - see the Configuration Reference.

3. Connect a StandIn identity

StandIn is the hosted service that joins the Teams call and dials into your bridge. Pick a tier at standin.komaa.com (sandbox for an instant trial), pair, and you get a shared secret.

1. Put the secret in WORKER_SHARED_SECRET (both sides must match exactly).
2. Point the identity’s agent WebSocket URL at your bridge, for example wss://dg-bridge.example.com:8080/voice/msteams/stream. StandIn appends /{callId} per call.
3. Restart the bridge if you changed the env.

StandIn dials in from the internet, so a laptop or private host needs a public URL. A tunnel gives you one and terminates TLS (so you get wss:// for free). Run one pointing at port 8080, then use the wss://…/voice/msteams/stream form of the printed host:

Tailscale Funnel:

tailscale funnel --bg --https=8080 8080

Cloudflare Tunnel:

cloudflared tunnel --url http://localhost:8080

ngrok:

ngrok http 8080

VS Code dev tunnels:

devtunnel host -p 8080 --allow-anonymous

For a fixed production host use an ingress/load balancer, or serve TLS natively with TLS_CERT_PATH + TLS_KEY_PATH. Never give StandIn a plain ws:// URL outside local testing.

More detail (tiers, what pairing does, cutoff behavior): Connecting to StandIn.

4. Make the first call

Call your Teams bot (or join the sandbox meeting). In the bridge logs you should see the call arrive, the Voice Agent session open, and the relay start:

INFO  [server] worker connected for call 19:meeting_ab... (1/64)
INFO  [call:19:meeting_ab] session.start (direction=inbound, recording=unknown)
INFO  [call:19:meeting_ab] Deepgram Voice Agent session open; waiting for SettingsApplied
INFO  [call:19:meeting_ab] SettingsApplied; relaying

Speak, and the agent answers in its own voice. If the call connects but something is off, Troubleshooting maps every error you are likely to see (401 handshake, agent-unavailable, a missing think endpoint) to its cause.