Developer Tools for Building AI Agents on Meeting Data
The layers you build on to put an AI agent on top of live and recorded meetings, and how MeetStream gives you the capture and real-time voice foundation in one API.
Meeting data is one of the richest inputs an AI agent can act on: who said what, when, and in what tone, across Zoom, Google Meet, and Microsoft Teams. But the meeting platforms were not designed to hand that data to your code, so you need an infrastructure layer that captures it and a set of tools you build on top.
Those tools fall into distinct layers, and mixing them up is the most common mistake teams make. The bot that joins the call and captures the media is not the same thing as the engine that turns audio into text, which is not the same thing as the framework that lets your agent talk back.
MeetStream is the capture and real-time voice layer: one API that joins the meeting, returns speaker-identified audio, video, and transcripts, and streams voice both ways. This guide explains each layer, then shows the speech engines and voice-agent frameworks you pair with MeetStream to complete the stack.
The tools and APIs
| Tool | Category | What it does | Best for |
|---|---|---|---|
| MeetStream | Bot/capture API | One API deploys a bot into Zoom, Google Meet, and Microsoft Teams to record, transcribe, and stream data. It returns speaker-identified transcripts, per-participant audio and video, and real-time audio over WebSocket as binary PCM frames tagged with speaker id. | Developers who want meeting-bot infrastructure with built-in real-time voice and streaming, without stitching a bot, an STT engine, and a media pipeline together themselves. |
| Deepgram | Speech-to-text | A speech recognition API with streaming and pre-recorded models. It converts raw audio to text but does not join meetings or capture the audio for you. | Streaming, low-latency transcription when you already have an audio stream and want to optimize for cost at volume. |
| AssemblyAI | Speech-to-text | A speech-to-text API with streaming and post-call models plus features like diarization and entity detection. Like Deepgram, it needs an audio source; it does not capture meetings itself. | High-accuracy post-call transcripts and voice-agent audio pipelines that already have raw audio in hand. |
| LiveKit | Voice-agent framework | An open-source WebRTC framework where your agent joins a room as a participant, subscribes to microphone tracks, and publishes audio back. It handles media transport, not meeting-platform bots. | Building a real-time voice agent that speaks and listens inside your own app or a WebRTC room. |
| Pipecat | Voice-agent framework | An open-source Python framework that orchestrates the voice loop of VAD, STT, LLM, and TTS with interruption handling. It is transport-agnostic and often paired with a media layer. | Developers who want fine-grained control over the pieces of a voice agent pipeline. |
| Vapi | Voice-agent framework | A managed voice-agent platform where you configure STT, LLM, and TTS providers and it runs the listen-think-speak loop for you. | Prototypes and simpler voice agents where you want the orchestration handled and less DevOps. |
The four layers, and why the distinction matters
Almost every tool in this space sits in one of four layers. Bot capture APIs put a participant in the call and hand you the raw media. Speech-to-text engines turn audio into text. Voice-agent frameworks let an agent listen and talk in real time. Notetaker products are finished apps that do the whole job for an end user.
The layers compose. A meeting intelligence product often uses a bot capture API for the media, a speech-to-text engine for the transcript, and a voice-agent framework if the agent needs to speak back. Knowing which layer you are missing is the fastest way to shortlist the right tool.
- Bot capture API: joins Zoom, Meet, and Teams, returns audio, video, and transcripts (MeetStream)
- Speech-to-text: converts audio to text but needs an audio source (Deepgram, AssemblyAI)
- Voice-agent framework: real-time listen and speak loop (LiveKit, Pipecat, Vapi)
How to evaluate a tool for your build
Start from what your agent needs to do, not from a feature list. If it only reads finished transcripts, you may not need a bot at all. If it acts during the call, you need real-time capture and low latency. If it talks back, you need a voice-agent layer on top of the media.
- Platform coverage: does it support all of Zoom, Google Meet, and Microsoft Teams
- Real-time vs post-call: do you need data mid-meeting or only the authoritative transcript afterward
- Per-participant streams: can you get audio and video split by speaker, not just a mixed track
- Latency: how many milliseconds from spoken word to your server
- Data control: can you bring your own storage, and what are the compliance credentials
- SDK and language support: how quickly can your team integrate it
Real-time versus post-call
Streaming transcription gives you words within roughly a second or two, which is what live coaching, real-time prompts, and live captions require. Post-call transcription is usually more accurate because the engine has the full audio context and can do better language-model rescoring.
Most meeting products want both: a streaming path for in-call features and a post-call path for the record of truth. A capture layer that supports both real-time and post-call transcription saves you from running two separate integrations.
Per-participant audio and video
A single mixed track is enough for a basic transcript, but agents that reason about who said what need separate streams. Per-participant audio lets you attribute speech precisely and run speaker-specific analysis. Per-participant video opens up presence, engagement, and visual context.
This is one of the harder things to build yourself, because conferencing platforms do not expose clean per-speaker streams. It is worth checking early whether a capture API returns separate streams or only a mix.
Build versus buy
The hard part of a meeting bot is not the model prompt. It is capturing raw audio from platforms whose APIs were not designed for continuous streaming, splitting the stream by speaker in real time, and keeping latency low, all while the bot stays reliably in the call.
Teams that build this in-house typically spend months on the media pipeline, hosting, and permissions work before they write a line of product logic. Buying a capture API collapses that to an integration you can stand up in hours. Building makes sense when data sovereignty or deep platform control outweighs speed, and open-source options exist for that path.
How MeetStream fits in
MeetStream is the meeting-bot-infrastructure layer with real-time voice built in. A single POST to https://api.meetstream.ai/api/v1/bots/create_bot, with a meeting link and a bot name, deploys a bot into Zoom, Google Meet, or Microsoft Teams.
From that one bot you get recording, speaker-identified transcripts, per-participant audio and video, real-time audio over WebSocket at roughly 200ms as binary PCM frames tagged with speaker id, real-time and post-call transcription, webhooks, MIA voice agents, calendar auto-join, and bring-your-own S3 storage. SDKs cover eight languages, and the platform carries ISO 27001 and GDPR compliance with SOC 2 Type 2 under audit.
In stack terms, MeetStream covers the capture layer and the streaming voice layer together, so you can pair it with your own choice of LLM and, if you want, an external speech engine, without owning the meeting-platform plumbing yourself. Pricing starts at $0.35/hr, drops to $0.25/hr at volume, adds $0.10/hr for transcription, and includes $5 in free credit to start.
Frequently asked questions
What is the difference between a meeting bot API and a speech-to-text API?
A meeting bot API joins the call as a participant and captures the raw audio, video, and metadata for you. A speech-to-text API only converts audio you already have into text. Most meeting products need both: a bot to capture the media and an engine to transcribe it.
How do I get real-time audio from a Zoom, Google Meet, or Teams meeting?
Use a bot capture API that deploys a participant into the call and streams the audio to your server over a WebSocket. With MeetStream, one API call returns real-time audio as binary PCM frames tagged with speaker id at roughly 200ms, across all three platforms.
Can I build my own AI notetaker on this stack?
Yes. A notetaker is an app built on a capture-and-transcription layer plus your own summary logic. MeetStream gives you the recording, per-participant audio, and speaker-identified transcript from one API, and you add the summary and action-item generation on top with your own model.
Is streaming or post-call transcription more accurate?
Post-call transcription is usually more accurate because the engine has the full audio context and can rescore. Streaming trades some accuracy for speed, giving you words within a second or two. Live features need streaming; the authoritative record is best taken post-call.
Do I need a voice-agent framework like LiveKit or Pipecat?
Only if your agent has to speak back inside the meeting. Frameworks like LiveKit, Pipecat, and Vapi run the listen-think-speak loop. If your agent only reads and analyzes meeting data, a capture API and a model are enough, and MeetStream includes voice agents natively.
How do I get audio separated by speaker?
You need a capture layer that returns per-participant streams rather than a single mixed track. MeetStream provides per-participant audio and video plus speaker-identified transcripts, so you can attribute every line to the person who spoke it.
Build it on MeetStream
One API to join, record, stream, and transcribe meetings across Zoom, Google Meet, and Microsoft Teams.