Privacy
Where your words actually go.
Most privacy policies are written so that nothing in them can ever be wrong. This one is written so you can check it. Where a third party sees your data, it is named. Where something leaves your machine, it says so. Where we say we do better than Google Meet or Zoom, their own pages are linked beside the claim.
Last updated 12 October 2026.
The short version
- The call itself never touches our servers. Video and audio go browser to browser, encrypted between the people in the call. When two browsers cannot reach each other directly, the call goes through a relay that passes along encrypted packets it cannot open.
- What does reach us is what the product is made of: the transcript, the chat, and the recording. Every meeting is recorded, so it lands in its host's dashboard — and anyone can leave themselves out.
- Anyone can be left out of the recording, the transcript and every AI, without leaving the call. Their picture and voice are not recorded, nothing they say is written down, and no AI reads them. The room sees that they asked.
- We learn nothing about your calls unless you turn on Help improve TeamMeet. It is off until you do.
- Nobody trains a model on your meetings — not us, and not the companies we send work to. Anthropic does not train on API data, and every transcription stream we send to Deepgram opts out of its model-improvement program.
- No ads, no tracking cookies, no third-party analytics. We count visits to these pages ourselves — which page, the site that sent you, and any campaign tag in the address — as daily totals, with no cookie, no IP address and nothing that tells one visitor from another.
How this compares to Google Meet and Zoom
Both are serious about security, and on some things — certifications, controls for very large organisations — they are ahead of us. On the things below we chose differently. Every claim about them quotes or links their own page.
| TeamMeet | Google Meet | Zoom | |
|---|---|---|---|
| Staying in the call but out of the recording and the AI notes | Anyone, any time, for themselves: picture, voice, words and every AI. Their tile says “not recorded” to everyone. | Consent is an admin setting, and a decline has two outcomes, both the admin's choice: “If a participant declines consent, you can either require their removal from the meeting” — or let them stay by stopping the recording, transcript or notes for everyone. support.google.com · admin help | Opting out of a recording means leaving: “Click Leave and then Leave Meeting to opt out and exit the meeting.” Only the host and co-host can stop the AI meeting summary. support.zoom.com · AI summary |
| What they learn about your calls | Nothing, unless you opt in. Then one anonymous call-quality report per call, made only of the counts listed below. No IP address is kept. | Google collects “Service Data”, including device identifiers and IP addresses, and uses it to improve its services. We found no switch that turns it off. cloud.google.com | Zoom owns the “Service Generated Data” it collects, and its “Required Diagnostic Data setting cannot be disabled.” zoom.com terms · diagnostic data |
| Who can open a meeting's transcript and recording | The host, and whoever the host gives a link to or shares it with — on purpose, one meeting at a time. No administrator account, ours or your company's, can open everyone's meetings. | Transcripts are saved in the organizer's Drive: “All invitees in the host's organization can open the attachment.” Admins can search and export Meet data with Vault. support.google.com · Vault | “Account owners can view and share recordings and transcripts of meetings” held on their account. zoom.com privacy statement |
| Keeping the call itself off the provider's servers | Always — calls are peer to peer by design, and transcription, notes and recording still work. (Transcription sends each person's own microphone to our transcriber; see below.) | Client-side encryption needs an Enterprise Plus or Education plan and an outside identity provider, and turns off recordings, transcripts, captions and “Take notes for me”. support.google.com | End-to-end encryption is off by default; turning it on disables the AI features, cloud recording, live transcription and the web app. support.zoom.com |
| Training AI on your meetings | No — and the companies we send work to are held to the same: Anthropic does not train on API data, and Deepgram is told not to keep your audio to improve its models. | Not without the customer's permission. workspace terms | No: Zoom “does not use any of your audio, video, chat…” to train models. zoom.com terms |
Checked against the linked pages on 12 October 2026; the Google Meet consent row again on 25 October 2026, when Google's admin page had changed its wording. If one of them has changed, write to us and this table changes. The same comparison with Microsoft Teams as well is on Why TeamMeet.
Leave me out
In a call: Settings (the cog) → Your privacy → Leave me out of the recording, the transcript and the AI notes. It is remembered on this device, so if it is on before you join it applies from your first second in the room.
- Your picture and voice are not recorded. Every recorder in the room stops drawing your tile and unplugs your voice from its sound within half a second. If you share your screen, that is not recorded either.
- Nothing you say is written down. The server refuses your lines — it does not rely on your browser to stop sending them — so you are not in the transcript, the live notes, the recap, the action items, the workspace, Ova's answers, Counsel's hearing in a Legal call, or anything an agent or a webhook receives.
- You are still in the call. People see and hear you live, as before, and they still see what you type in the chat. Agents and webhooks are not sent your chat.
- Asking is still yours. If you put a question to Ova yourself, or ask the workspace or a drawing for something, that request is handled like anyone's — you chose to ask it, in front of the room.
- The room is told. Your tile carries a “not recorded” mark on everyone's screen, and the chat says you asked. A person missing from the record should never be a surprise to the people in it.
- It works from the moment you turn it on. It does not erase what was recorded before, and it cannot stop the people still being recorded from repeating what you said.
Help improve TeamMeet — off unless you turn it on
In a call: Settings → Your privacy → Help improve TeamMeet. It is off by default, and it is per device. When it is on, your browser sends one report when a call ends, and this is everything that can be in it:
- how many minutes the call lasted
- how many people were in it
- phone, tablet or computer
- the browser and the operating system, by family
- whether a relay was available
- how many connections were tried, got through, went through the relay, failed, or were dialled again
- how many times the connection to us dropped
- how many times your camera had to be restarted
- how many times your browser blocked the sound
- which layouts you used
Never: names, the meeting's link, anyone's id, words, pictures or sound. The server keeps only the fields above — anything else in a report is discarded — with the hour it arrived rather than the second, and without the IP address it came from. Reports are deleted after 90 days. They exist to answer one question: how often does a call fail to connect somebody, and on what.
When you send feedback yourself, the feedback form adds the shape of the room it was sent from — the layout, how many people, the screen size, the browser — because a bug in a call cannot be fixed without it. You see the form, and you press send.
So we can tell you when it ships, we keep with it how to reach you: the email address you gave or your account's, and a scrambled form of your browser's id — never shown on the public page for it. When it ships we send one email, show a notice in any call you are in, and show one on your next visit, once.
What we hold, and for how long
| What | Where it lives | How long |
|---|---|---|
| Your account, if you sign in with Google | Your name, email address and profile picture, on our disk | Until you ask us to delete it. Everything works without signing in. |
| A meeting while it runs | Names, the chat and the transcript, in server memory — written to our disk for the few minutes a restart takes | About four hours after the last person leaves |
| Recordings | The video, its transcript, the recap and the action items, on our disk — owned by the host's account, or by whoever holds the owner key for a guest | Until the owner deletes it |
| Files dropped into the chat | Our disk | Until deleted. Anyone with the link can fetch them |
| Pictures drawn in a meeting | Our disk | Until deleted |
| A room's workspace board | Our disk | Until deleted. Readable by people who own a recording made in that room |
| Feedback and fix requests | Our disk. A fix request is public by its link — without your name or email | Until deleted |
| Call-quality reports | Our disk, only if you opted in, as anonymous counts | 90 days |
| Page visit counts | Our disk, as one set of totals a day: page views and first visits for each kind of page (never the address of a meeting, recording or booking), and for first visits the name of the site that linked here and the utm_ campaign tags. No cookie, no IP address, no browser details | Until deleted. A day’s totals cannot be traced back to anyone |
| Developer keys | Our disk, as a SHA-256 hash — we cannot read a key back | Until deleted |
| A session cookie, if you sign in | Your browser, signed so it cannot be forged | Until you sign out |
Your browser also keeps your own settings — your name, your language, your camera choice, the two privacy switches above, and a random id for this browser that lets the room recognise your old tab if it was closed without saying goodbye. They never leave your device except as the room needs them. Each tab also keeps, until it is closed, a note that it has already been counted as a first visit.
Transcription, honestly
Each microphone is transcribed on its own, which is why the speaker on every line is known rather than guessed. How that happens depends on what is switched on:
- Our transcriber (Deepgram Nova-2). Your browser streams your microphone to our server, which passes it to Deepgram and passes the words back. We do not store the audio. Every stream asks Deepgram not to keep it for model improvement, so Deepgram keeps it “only for the duration necessary to process the request” (their words).
- Chrome and Edge on a computer also run the browser's own speech recognition, for the quickest live captions. Chrome sends that audio to Google's speech servers, under Google's terms.
- When our transcriber is unavailable, Chrome and Edge use the browser's own recognition alone, and Safari and Firefox run a Whisper model inside the tab — nothing leaves your machine.
- Either way, the words come to our server, because the room is built out of them: the notes, the agenda, Ova and the search all read the same transcript.
- You can stop it. Settings → Transcription turns off writing down what you say. Leave me out does that and keeps you out of the recording and the AI too.
Who else sees it
Named, because “trusted third-party service providers” tells you nothing.
| Fly.io | Hosts the server and the disk. It sees the network address of every connection, to route it. |
| Anthropic | Transcript text, to write the live notes, the recap, the agenda, the workspace and Ova's answers — and, when someone in the call turns on the live story, the last few seconds of what people said, to draw it. Anthropic does not train its models on API data. |
| Deepgram | Your microphone's audio, streamed, to transcribe it — opted out of its model-improvement program, as above. |
| Audio, for the browser's own speech recognition in Chrome and Edge. Your name, email and picture, if you sign in with Google. A network address, for the STUN server that helps two browsers find each other. Our pages load their typefaces from Google Fonts, which sees your address too. | |
| Cloudflare | A network address, for its STUN server — and, once switched on, the relay that carries a call two browsers cannot connect directly. The relay passes encrypted packets it cannot open. |
| Fish Audio | The text of what Ova is about to say, to speak it. |
| fal.ai | An image prompt written from the conversation, when a picture is drawn. |
| Resend | An email address and the booking's details, to send a booking confirmation. |
| Recall.ai | Only when you send the notetaker to someone else's Zoom, Meet, Teams or Webex call: that call's link, and the recording and transcript its bot makes there. |
| OpenAI | Only in a Legal call with Counsel in it, once Counsel's voice is switched on: the call's audio and the document, for Counsel to hear and read. OpenAI does not train on API data by default. |
| Stripe | Only if you buy a plan: payment details, typed into Stripe's page — we never see the card. |
That list is complete as of the date at the top. If it changes, this page changes.
Teams
A team shares meetings by two explicit acts — somebody invites you, and somebody shares a meeting into it. Being in a room once, or having your name in somebody's transcript, gives you nothing. When a person leaves a team or a meeting is un-shared, the access is deleted, not flagged.
Agents
A program can join a meeting with a key. Everything about how we did that is a response to how badly it goes elsewhere:
- An agent is in the member list, like a person. There is no hidden listener.
- Its arrival is written into the chat, where everyone sees it.
- Anyone in the room can remove it — not just the host.
- The host can turn agents off entirely with one switch, and every connected agent is disconnected and told why.
- An agent receives the transcript and the chat. It does not receive raw audio or video — and from anyone who asked to be left out it receives neither their words nor their chat, only a question they chose to put to Ova, which the whole room sees.
- Someone can connect their own Claude and have the call hand it a task. The task carries the words it was asked in and the lines said just before it — never a line from anyone who asked to be left out — and it goes to that person’s own Claude account, which the whole call is told in the chat. Their Claude can only join a call they are in themselves.
Recording
Meetings record themselves — once a second person arrives, or once anyone is transcribed — so the host finds them in their dashboard without having pressed anything. One person's browser makes the recording — it draws the tiles and mixes the sound — and uploads it to us in pieces while the meeting runs. Everyone sees the recording indicator for as long as it runs, and anyone can leave themselves out of it. There is no silent recording.
A recording is owned by the host's account and is either private — only its owner can open it — or open to anyone with its link. Nothing is ever listed publicly. Laws about recording people differ by where everyone is sitting, and complying with them is the responsibility of the person who runs the meeting.
What we do not do
- We do not train models on your meetings, and we hold the companies we send work to to the same.
- We do not sell or rent your data.
- We do not run advertising, and nobody pays us to put anything in front of you.
- We do not use tracking cookies or third-party analytics.
- We do not read your meetings. The transcript is processed by software to build the notes; nobody here opens a meeting unless its owner asks us to, or the law makes us.
- We do not collect anything about your calls to improve the product unless you have turned that on yourself.
Your choices
- Leave yourself out of the recording, the transcript and the AI, in any call.
- Stop transcription of your own voice.
- Delete your recordings from their page or your dashboard. Deleting removes the video, its transcript and its notes from our disk.
- Opt in, or back out, of call-quality reports at any time.
- Ask us to delete your account, or anything else we hold, at the address below.
If you are in the EU or UK, our legal basis for processing meeting content is legitimate interest in providing a meeting you asked to be in; for call-quality reports it is your consent, which you can withdraw by switching them off.
Children
Team Meet is not directed at children under 13, and we do not knowingly collect their information. A school using this should be the one deciding whether that is appropriate, with consent from the people it needs consent from.
Contact
Questions, deletions, or a claim that something on this page is wrong: privacy@teammeet.io. The last one is the most useful kind of mail we can get.