Skip to main content
@komaa/openai-msteams-bridge brings an OpenAI Realtime voice agent (gpt-realtime - speech-to-speech: STT + LLM + TTS + turn-taking in one model) onto a Microsoft Teams call. It is the OpenAI analogue of the ElevenLabs and LiveKit bridges: it hosts the HMAC WebSocket that StandIn connects to, then relays audio to and from an OpenAI Realtime session.
Like the other bridges, this is a standalone Node service, not a framework plugin. There is also no agent to configure in a dashboard: the bridge configures each Realtime session itself (model, voice, instructions, VAD, tools) from environment variables. Two variables and it runs.

​
How it works

The bridge opens one Realtime WebSocket per call (wss://api.openai.com/v1/realtime), configures the session with your instructions plus per-call caller context, and relays audio both ways. The Teams wire is PCM 16 kHz; the Realtime API speaks PCM 24 kHz only, so the bridge converts at that one boundary with a built-in anti-aliased resampler (no codecs, no external dependencies). Caller barge-in maps onto the model’s own interruption handling, and stale “ghost” audio is dropped so nothing plays after the caller cuts in. The agent gets four built-in tools automatically (end_call, look for vision, show_image, express for avatar emotion) - and you can register your own function tools or pass remote MCP servers the Realtime API executes server-side. See Extending the agent’s tools.

​
Requirements

  • Node.js >= 20.
  • An OpenAI API key with Realtime access (platform.openai.com).
  • A StandIn subscription (standin.komaa.com, free package works) - the hosted media bridge that joins the Teams call and connects to this service. See Architecture.
  • Your own Microsoft Teams bot connected to StandIn - the Teams setup walks through the Azure bot, the app package, and the upload. To try it without one, use the sandbox.

​
Run

Env-configured, no install step needed: 
OPENAI_API_KEY=sk-... \
WORKER_SHARED_SECRET=... \
npx @komaa/openai-msteams-bridge
Optionally shape the agent (all have defaults): 
OPENAI_VOICE=marin
OPENAI_INSTRUCTIONS="You are Komaa's friendly receptionist. Keep replies short."
OPENAI_FIRST_MESSAGE="Hello! You've reached Komaa. How can I help?"
Or add it to a project (npm i @komaa/openai-msteams-bridge) and embed it - see the library API for the programmatic surface, custom tools, and a custom vision hook. The bridge binds its media WebSocket (default ws://<host>:8080/voice/msteams/stream; StandIn appends /{callId} per call).
The OpenAI Realtime API caps a session at 60 minutes server-side. Set MAX_CALL_MINUTES below 60 so the bridge’s own governor ends long calls with a spoken goodbye instead of an abrupt drop - the bridge warns at startup when it would not.

​
Expose the WebSocket (Agent voice URL)

StandIn connects to the bridge from the internet, so the port must be reachable - a public host or a tunnel. With Tailscale Funnel: 
tailscale funnel --bg --https=8080 8080
Your Agent voice URL is then: 
wss://<machine>.<tailnet>.ts.net:8080/voice/msteams/stream

​
Connect it to StandIn

Register that URL (with a matching shared secret) on your identity in the dashboard, or use the sandbox to try it without your own Teams bot:
  1. Set the identity’s Agent voice URL to where the bridge listens.
  2. Set WORKER_SHARED_SECRET to the shared secret from pairing - both sides must match exactly, or the WebSocket handshake is rejected with 401.
  3. Place a Teams call (or join the sandbox meeting). StandIn joins, connects to the bridge, and your Realtime agent answers.

​
Try the runnable example

The fastest way to understand the bridge is to run its example project - a minimal, working embedding you can copy straight into your own repo.

Run the example, step by step

Clone examples/basic-bridge, fill in two environment variables, expose the port, connect it to StandIn, and place a call - with the vision hook and a custom lookup_order tool explained along the way.
Next: the full configuration reference - every environment variable, the call governor, vision, and MCP servers. Deep protocol and library docs live on the project site.

Docs site

npm

Source