Mainroom FAQ

Mainroom is a platform for building AI agents that join live meetings on Google Meet, Microsoft Teams and Zoom as participants. An agent has a camera tile and a voice, listens to the whole room, speaks when addressed, presents on the shared screen, draws on a whiteboard, runs live browser demos, looks things up in connected tools, and emails a recap afterwards. You invite one by adding its email address to a calendar invite.

Google Meet, Microsoft Teams and Zoom. The agent joins as a regular participant and is admitted from the lobby like any guest. For the full experience, enable captions (that is how it knows who is speaking by name) and allow participants to share their screen. On Google Meet, hosts may find the agent in a second list of join requests that Meet holds back and sets to Deny by default (introduced in spring 2026); if it is not in the main list, admit it from there.

Add the agent's email address to the calendar invite, or forward the invite to it. The agent joins at the start time and appears in the lobby; the host admits it. There is nothing to install on the meeting platform. For a recurring meeting, invite it to the series once: it joins every occurrence, skips the ones you cancel, and follows moved instances.

Yes. An agent can run a meeting, a class or a practice session in another language: give it a persona that speaks that language. On Google Meet its captions follow the agent's language, and when people switch languages mid-call the agent keeps its own transcript of what was said, so the recap stays readable.

No. An AI notetaker records and summarizes. A Mainroom agent participates: it answers questions in the call, puts material on the shared screen, runs demos, teaches on a whiteboard, gathers input from the room, and drafts follow-ups. It also produces a recap, but the recap is a by-product rather than the product.

No. It speaks when spoken to, waits for a gap before reporting findings, and stops the moment a person talks over it. Tell it to go on mute and it stays silent until someone asks it to come off mute. You write its meeting manners in plain language, and the customer-facing templates default to silent unless named.

You choose per agent: only you, or anyone with your company's email domain, with no Mainroom account required for colleagues. Agents you build stay yours.

Whoever invited the agent gets a link to my.mainroom.ai the first time it joins a call. During the call it shows live indicators (talk share, questions asked, longest stretch), short private nudges, and a place to ask the agent questions quietly. Nothing on it appears in the meeting.

Agents announce themselves in the roster and platforms show their standard recording notice. The transcript powers the recap email and, if enabled, the agent's memory; memory and recaps are per-agent switches. Credentials for connected tools are stored server-side and never exposed to the meeting. See the privacy policy for details.

Only as a proposal. Anything that changes a system of record (a CRM note, a ticket, a follow-up email) arrives as a proposed action that the agent's owner approves or discards. Nothing is written without approval.

Any tool that exposes an MCP server (Model Context Protocol), plus built-in web search. Connections are configured per agent in the Studio.

Tell it to leave and it says goodbye and leaves. You can also remove it from the invite, or have the host remove it from the call like any participant. Everyone in the room can see it is there; it never joins silently.

Mainroom is free while in beta. Build agents, invite them to real meetings, and tell us what you would pay for.

Yes. Mainroom exposes a public MCP server at mcp.mainroom.ai and an A2A agent (Agent Card at mcp.mainroom.ai/.well-known/agent-card.json) with skills to describe the product, list agent templates, and send the demo agent into a meeting; a Gemini Enterprise admin can register it as a custom A2A agent. This site also publishes llms.txt and llms-full.txt, and every page has a markdown twin at the same path with .md appended.