MeetStream AI

Unified API & Infra for AI Agents in Meetings

Visit Website
September 6, 2026 1M meeting minutes processed. What we learned about how developers actually onboard

MeetStream crossed one million meeting minutes processed last week, across Zoom, Google Meet and Microsoft Teams. This post is about what we learned watching developers go from signup to their first bot in a meeting, and what changed in the product because of it.

What we are

One API to put a bot into any meeting. It joins, records, streams every participant's audio and video as separate tracks, and pushes the live transcript over a websocket while the call is still running. Voice agents can listen and speak back inside the meeting. There is an MCP server and a CLI so agent runtimes can drive it directly.

The user journey we see most often

1. Someone signs up, usually from a search like "Google Meet bot API" or from a comparison post.

2. They copy the create-bot request from the docs, paste their own meeting link, and send it. The bot shows up in their call within about a minute.

3. They watch the transcript stream in, then go looking for per-participant audio. That is usually the moment they decide to stay.

4. The first real integration is almost always a note taker or a CRM updater. Voice agents come second, once they trust the audio path.

Half of all signups run a bot, and most of those do it the same day they sign up. Those two numbers are the ones we watch most.

What we changed because of that

- Free credit with no card, so step 2 happens without a sales call.

- The docs open on a working request, not an overview page. Time to first bot went from hours to minutes.

- Per-participant audio and video on every platform, because it was the feature people asked about in the first ten minutes.

- Transcription re-trigger, bring-your-own S3, and a choice of transcription providers, because the second week is when people start caring about pipelines and data ownership.

- An MCP server, because a lot of new users found us through Claude or Cursor before they found the website.

What did not work

- A big coordinated launch on a date. It slipped. What moved signups was shipping and writing every week.

- Generic overview content. Posts that answer one specific developer question outperformed everything else.

- Making claims early. We rounded up "one million minutes" before it was true. It became true on September 1. Do not round up.

What we would do differently

Build the self-serve path first and everything else second. Our first months were the reverse, and we had plenty of interest that never turned into a bot in a meeting.

What is next

More voice agent tooling is on the way, and we are spending time with the teams running the most minutes to see what they need next. Happy to answer anything about meeting bot infrastructure or agents that need to be in a call.

Comment

About

Every AI product will eventually need to be in the room. We built the infrastructure that puts an agent into any meeting with one API call, hears every participant, and talks back.