AI Voice Agents That Join, Speak, and Act in Live Meetings

What are AI voice agents in meetings?

An AI voice agent is a program that joins a live meeting as a participant, listens to what people say, and responds out loud in real time. It is not a notetaker that summarizes the call afterward, and it is not a phone or web voice bot that lives outside your meetings. It sits inside the Zoom, Google Meet, or Microsoft Teams call and takes part in the conversation.

That distinction decides what you can build. A passive bot can transcribe a sales call; a voice agent can answer a prospect's question mid-pitch. A notetaker can log an interview; a voice agent can run the interview. Once an agent can hear per-speaker audio and speak back into the room, the meeting becomes something your software takes part in rather than something it only records.

This page covers how these agents work, what you can build, and why MeetStream is the first meeting bot API to make voice a built-in part of the agent instead of a stack you assemble yourself.

How voice agents work: STT, LLM, and TTS

A voice agent runs a loop. Speech-to-text (STT) turns what a speaker says into text. A large language model (LLM) reasons over that text and decides what to say. Text-to-speech (TTS) turns the reply into audio that plays into the meeting. The loop repeats every turn.

The hard part is time. Human conversation leaves a window of roughly 300 to 500 milliseconds before a pause feels like a stall. To stay inside it, each stage streams into the next rather than waiting for the one before it to finish, so the agent starts forming a reply while the speaker is still talking. Latency, not raw accuracy, is usually what decides whether an agent feels present or awkward.

There are two ways to assemble that loop: a single realtime model that goes speech to speech in one pass, or a pipeline of separate STT, LLM, and TTS providers you choose. MeetStream supports both, and we compare them below.

How MeetStream and MIA deploy voice into a live meeting

MeetStream deploys voice agents through MIA, the MeetStream Infrastructure Agent. One API call sends a bot into a Zoom, Google Meet, or Microsoft Teams meeting, opens per-participant audio so the agent hears each speaker cleanly, and streams the agent's synthesized voice back into the call. The same bot that records and transcribes is the one that speaks.

That is what built-in voice infrastructure means. You do not run a recording bot on one system and a voice stack on another and try to keep them in sync. You configure an agent, attach it to a bot with an agent_config_id, and audio in and audio out are handled on the same connection.

The shape of it: create a bot, point it at a meeting, and give it an agent to run.

$ curl -X POST https://api.meetstream.ai/api/v1/bots/create_bot \
    -H "Authorization: Token $API_KEY" \
    -d '{
      "meeting_link": "https://meet.google.com/abc-defg-hij",
      "bot_name": "Voice Agent",
      "agent_config_id": "agent_your_mia_config",
      "live_audio_required": { "websocket_url": "wss://your-app/audio" }
    }'

Realtime mode vs pipeline mode

MeetStream runs agents in two modes. Realtime mode uses one low-latency speech-to-speech provider for the fastest turn-taking, which suits agents that need to feel instant. Pipeline mode composes separate STT, LLM, and TTS providers, so you pick the transcription engine, the reasoning model, and the voice independently.

The tradeoff is latency against control. Realtime mode is the shortest path to a natural-feeling conversation. Pipeline mode gives you model choice at every step, which matters when you need a specific reasoning model, a particular voice, or a transcription provider tuned to your vocabulary. You choose per agent, so a support agent and a sales agent on the same account can run different stacks.

Agents that take actions with MCP tools

An agent that only talks is a demo. An agent that can act is a product. MeetStream agents call external tools during the conversation through the Model Context Protocol (MCP), the open standard introduced by Anthropic for connecting agents to tools and data.

In practice, an in-meeting agent can look up the account being discussed, update a CRM record, create a ticket, or pull an answer from your knowledge base while the call is live, then reflect the result back in voice and in the meeting chat. Its response types are voice, chat, and action, so it can speak, post, and do in the same turn.

This is what separates a voice agent from a voice demo. It reaches into the systems your team already runs and changes something, rather than only describing what it would do.

Why MeetStream is the first built-in voice infrastructure for meeting agents

Most tools in this space solve a neighboring problem. Recording-first meeting APIs added the ability to play audio into a call as a single feature, but the actual voice conversation is a stack you assemble yourself by wiring in a separate realtime model. Voice frameworks built for phone and web are strong at telephony and browser widgets, but they were not designed to join a scheduled Zoom, Google Meet, or Teams call as a named participant.

MeetStream is the first to treat voice as native infrastructure for an agent that joins the real meeting. The bot that joins, the per-participant audio it hears, and the voice it speaks are one system behind one API. We can claim first here because no other developer platform ships voice-in-a-real-meeting as a built-in primitive rather than a self-assembled integration.

The timing is not an accident. Zoom shipped Realtime Media Streams in 2025, Google Meet opened a Media API in developer preview, and the industry moved toward agents that stand in for people in meetings. The infrastructure to speak in a meeting is arriving, and MeetStream packages it so developers build on it directly.

Voice agents across Zoom, Google Meet, and Microsoft Teams

One integration covers all three major platforms. The same agent configuration that joins a Google Meet call joins a Zoom or Microsoft Teams call, with per-participant audio and video and voice output, and you do not branch your code per platform.

Where a platform exposes native real-time media, MeetStream uses it; where it does not, the bot still joins as a participant and speaks into the room. From your side the interface is identical, so you build the agent once and it works wherever your users meet.

Use cases: sales coaching, interviews, standups, and stand-in agents

The same primitive supports very different products. A sales agent can pitch and answer objections live on a discovery call. A recruiting agent can run a structured first-round interview and score it. A scrum agent can run a daily standup, ask each person for their update, and post the summary. A stand-in clone can cover a low-value meeting you would otherwise skip and brief you afterward.

What these share is presence. The agent is in the room, hearing each speaker and responding in the moment, which is only possible when voice is part of the meeting bot rather than bolted on beside it.

Start building voice agents

Voice agents in meetings stopped being a research demo and became something you can ship. The stack is real: per-participant audio in, a fast STT, LLM, and TTS loop, voice out, and tool actions through MCP, all on one API that joins Zoom, Google Meet, and Microsoft Teams.

MeetStream is the first to make that voice infrastructure built in rather than assembled. Point a bot at a meeting, attach an agent, and it joins, listens, speaks, and acts. Get started free at meetstream.ai, or see the full API reference at docs.meetstream.ai.

Frequently asked questions