Udumbara privacy policy
Last updated: 30 October 2026.
Udumbara is a Chrome extension made by SnapAI Solutions. It helps you during Google Meet calls (and Microsoft Teams calls in Chrome, if you switch that on) by suggesting questions, drafting answers from your documents and keeping a record of your meetings. This policy explains what it handles and where that goes.
The short version
- What Udumbara keeps is saved in your own OneDrive or Google Drive, in a folder only Udumbara uses, and in your browser on your computer.
- Udumbara's AI service processes each request and keeps nothing: no database, no files, no logs of what was said.
- Normally Udumbara reads only the captions the meeting shows on screen. If you switch on "listen to the call audio" and press its icon, it also captures the call's audio to turn into text when captions are off — played back so you still hear the call, turned into text by a speech service, and never stored.
- We don't sell your data, use it for advertising, or use it to train AI models.
What Udumbara handles
| What | Where it comes from | Where it goes |
|---|---|---|
| Captions of your Meet calls, and of your Teams calls in Chrome if you switch "Teams on the web" on (the text the meeting shows on screen) | The Meet or Teams tab, while you're in a call with captions on | Your browser, your drive, and the AI service for one request at a time |
| The call's audio — only if you switch "When captions are off: listen to the call audio" on and press the Udumbara icon in the call | The meeting tab's sound, and your microphone, captured through Chrome, while you're listening to audio (Pause stops it) | Streamed to the speech service (Deepgram) to turn into text; the text becomes transcript lines in your browser and drive. The audio itself is never stored — not by Udumbara, not in your drive, not by the AI service |
| The meeting app's official transcript, after the call (Teams via Microsoft Graph; needs your admin's consent) | Your meeting app, read with your own sign-in — only where transcription was on and your account can read it | Your browser and drive: it fills stretches Udumbara missed and proposes real names for audio speakers, which you confirm. Not available for most meetings; absence is normal |
| A transcript file you import into an engagement in the workspace (a .txt, .vtt or .srt of a meeting you already had), and your confirmation that you have permission to upload and process it | The file you choose; nothing is read or sent until you press Save and analyse | Your browser and your drive, as an ended meeting with the file's name, a fingerprint (SHA-256) of its text and when you confirmed; its lines go to the AI service to find what was said, like a meeting's captions, one request at a time. The AI service keeps nothing |
| Meeting records: transcript, ledger items (facts, decisions, questions, actions, terms), summaries, and your review of them (when you confirmed or rejected an item, your note on it, an owner or due date you typed) | Made from the captions, and edited by you | Your browser and your drive |
| Meeting outputs you ask for after a meeting (a short or detailed summary, one with citations, a summary with action items, highlights, tasks) | Written by the AI service when you press Generate, from that meeting's own transcript lines and ledger items only (never another meeting's, never your private notes) | Your browser and your drive, with the meeting. The AI service keeps nothing |
| Each engagement's context.md and memory.md: its briefing and what has been learned, and anything you write under "My notes" in them | Written by the AI service after each meeting with that engagement (and when you choose Update now), from that engagement's findings and its meetings' titles, dates and summaries; your "My notes" sections are yours and are kept as you wrote them (the Briefing you write on an engagement's Overview, before its first meeting or later, is the "My notes" of context.md) | Your drive, in the engagement's folder, and your browser. What you write under "My notes" in these two files is sent to the AI service with each update, as corrections it must follow. During a call with that engagement, context.md and the current-state part of memory.md (with your "My notes" in them) also go with each suggestion and answer draft as background, never another engagement's; your private notes for a meeting (below) never are |
| An engagement's summary (where it stands: what's decided, still open, at risk and changed lately) | Written by the AI service when you press Summarise or Refresh on the engagement's Overview, from that engagement's findings, its meetings' titles, dates and summaries, and its context.md and current state; never another engagement's, never your private notes | Your browser, and summary.md in the engagement's folder in your drive. The AI service keeps nothing |
| Questions you ask about an engagement, and their answers | Typed by you in "Ask about this engagement" on its Overview; answered by the AI service from that engagement's findings, its meetings' titles, dates and summaries, its context.md and current state, a few excerpts of its documents and a few short passages of its transcripts that match the question (found in your browser), and your last few questions and answers there; never another engagement's, never your private notes or your About you | Your browser only: the last 20 questions and answers per engagement stay in its working copy on this computer (Clear removes them; on the website they are deleted with it when you sign out). Not saved to your drive and never kept by the AI service or SnapAI |
| The meeting's type (Discovery, Design review or Status), which chooses the checklist the suggested questions aim at | Chosen by you in the Udumbara panel (Discovery unless you change it) | Your drive, with the meeting |
| About you (needed to start using Udumbara): your role, your main goal in meetings, the kinds of meetings you use Udumbara for, your experience in the role, how much detail and what tone you prefer, and up to 600 characters on what you care about in calls (optional; started for you from your role, and yours to change) | Chosen and typed by you in setup or Settings › About you | Your drive (profile.md in Udumbara's folder) and your browser; and, if you use Udumbara's website, your choices (not what you wrote about what you care about) on your Udumbara account too (see "A Udumbara account"), so setup carries on in any browser. On the website, until your drive is connected, what you wrote waits in your browser. Sent to the AI service with each suggestion and answer draft, so they're pitched to your role, experience, depth and tone; it's used only for the angle, depth and wording of the help, never as facts about a client, and never with ledger updates, summaries or meeting outputs. Change it in Settings; deleting your account removes it there. If you set a different role for an engagement (and, optionally, up to 300 characters on what you care about there), it is kept in that engagement's file in your drive, and sent only with that engagement's suggestions and answer drafts |
| Your own checklists, if you change one in Settings: area names and the words that match them | Typed by you in the personal space | Your drive (Udumbara's app folder). Area names go to the AI service with suggestions and ledger updates; the words don't |
| Your private notes for a meeting | Typed by you in the Udumbara panel (you can add lines suggested from the meeting's own record; nothing is added unless you press Add) | Your browser and your drive, with the meeting. Only you see them; never sent to the AI service or to SnapAI |
| Words still being spoken | The captions or audio, before a line is finished | Shown in your panel for a few seconds, then replaced by the finished line. Never stored, saved to your drive or sent anywhere |
| Documents you add for answers | You choose them | Your drive. PDF, Word, PowerPoint, email (.eml) and web page (.html) files are sent once to the AI service to read their text (an email's attachments are named, not read); the text is saved back to your drive. Relevant excerpts are sent with a question to draft an answer, together with (since slice EM4a) a few of the same engagement's findings, passages of its meetings' saved summaries and short passages of its earlier meetings' transcripts that match the question, each labelled with its meeting, and (since slice EM4c) the findings you confirmed in that engagement's earlier meetings (its decisions, answered questions and facts, up to about 32,000 characters); the suggested questions get the top few passages too, and the confirmed findings of earlier meetings in full (in place of those same findings in the engagement's list of findings they already get). Never another engagement's, never your private notes. A question they don't cover is sent with no excerpts, so the AI can give a short general-knowledge answer, marked as not from your documents. Removing a document in Udumbara moves it, and the text saved from it, to your drive's recycle bin or trash. |
| Your name and email | Your Microsoft or Google sign-in | Your browser, to show who is connected; and, if you have a Udumbara account, that account (see below) |
| Sign-in tokens | Microsoft or Google | Your browser; sent with each request so the AI service can check it's you |
| Udumbara account tokens (not open yet) | Udumbara's account service, five minutes at a time | On the website, your browser's memory for that tab; in the extension, once it has joined the account, its own session (ended after 30 days unused or when it leaves) in the extension's storage on this computer. Sent with each request to the AI service, carrying only your random account number |
| Pilot measures: which suggested answers you put in the chat box or copied, your 👍/👎 on drafts and suggested questions, when the client document was saved, when you paused listening, how often you collapsed or closed Udumbara in the call, and how much of the meeting type's checklist was covered when the meeting ended (a percentage and counts, not the findings) | Your clicks in Udumbara's side panel, its panel in the call and the personal space | Your browser and your drive, with each meeting. Never sent to the AI service or to SnapAI |
The AI service
When the extension needs a suggestion, an answer draft, ledger items, topics for an engagement's findings, a summary, a meeting output you asked for, an engagement's summary, an answer to a question you asked about an engagement or a Word document, it sends what that one request needs (recent lines of the meeting, the open items of the same engagement, a few document excerpts, a few of the same engagement's findings, summary passages and earlier transcript passages that match (slices EM4a, EM4b), and the findings you confirmed in its earlier meetings (EM4c), and for suggestions and suggested answers the same engagement's context.md and current state; for an engagement's summary or a question about it, that engagement's findings, its meetings' titles, dates and summaries, its context.md and current state, and for a question also your words, your last few questions and answers there, and a few short passages of its transcripts; for suggestions and ledger items also the names of the meeting type's checklist areas, such as "Integrations": Udumbara's own words, or yours if you changed the list) to Udumbara's AI service. The service checks your sign-in with Microsoft or Google, and sends the request to Anthropic's Claude to produce the result. It returns the result to your browser and keeps nothing. Its logs record only which kind of request it was, whether it worked, how many tokens it used and what that cost, with a short code made from your account id so usage can be added up per account; never what was said, your name or your email. Since 30 October 2026 the same numbers (the kind of request, the model, tokens, cost, how long it took and that short code) are also kept in SnapAI's usage records, each day's totals for 13 months, so SnapAI can see what the service costs and keep it within budget. Named SnapAI staff see them in SnapAI's private console; only its owners and finance staff see which account a short code belongs to.
Anthropic processes these requests as SnapAI's provider under its commercial terms, which do not allow Anthropic to train its models on them. Anthropic deletes API requests and responses from its systems within 30 days, longer only where its usage policy or the law requires it (how long Anthropic keeps data).
One engagement's records are never sent with another engagement's requests.
When the question check is switched on in the AI service (it is off on the service beta users connect to), the AI service also sends the line about to get a suggested answer, and the three lines spoken just before it, to Cloudflare's Workers AI, which returns only how likely the line is a real question, so lines like "Can you hear me?" don't get an answer drafted. Nothing else is sent with it: no documents, notes, names or profile. Cloudflare already runs Udumbara's AI service and processes this as SnapAI's provider under its terms for Workers AI.
The speech service (only when you listen to the call audio)
When you turn audio listening on and press the icon, the call's audio (the other participants, mixed, plus your microphone) is streamed to Deepgram, which turns it into text in real time and returns the words. The audio is used only to transcribe and is not stored. Deepgram is SnapAI's speech provider; on the hosted service your call's speech passes through Deepgram, and the on-premises edition can run Deepgram inside your own network so nothing leaves it. Every connection opts out of Deepgram's Model Improvement Program, so Deepgram keeps the audio only as long as it takes to transcribe it and never uses it to train its models. If you never turn audio listening on, no audio is ever captured.
Your drive
Udumbara asks only for access to its own folder: OneDrive › Apps › Udumbara (Files.ReadWrite.AppFolder), or the files it creates in Google Drive (drive.file). It can't see the rest of your drive. It does read how much free space your drive has (when the drive shows it), to warn you before it fills up; that number stays in your browser. You can read, move or delete the files at any time. Disconnecting in the extension stops new saves. Deleting the folder removes Udumbara's records from your drive.
A Udumbara account (for the website; not open yet)
When signing in on Udumbara's website opens (first to invited testers), SnapAI keeps a small account record so you can sign in with Google or Microsoft whichever drive you use. It holds only: a random account number; which Google or Microsoft sign-ins are linked to it (the provider's own id for you, not your email address); your signed-in browsers (a fingerprint of each, ended after 30 days without use or when you sign out); which version of the tester agreement and this policy you accepted and when; which drive your workspace uses (OneDrive or Google Drive, and that drive's own id for your account, not its address), so a new browser connects the same one; and (since 30 October 2026) the name and email address your latest Google or Microsoft sign-in gave, and your About you choices with when you saved them (since 30 October 2026, never what you wrote about what you care about in calls: that stays in your drive, and any such text the account held before was removed). Since 30 October 2026 accounts are made only with an invitation code SnapAI sends you (bound to your email address unless SnapAI chose otherwise): SnapAI keeps a fingerprint of each code (never the code itself), the group of testers it is for, the email address it was sent to (until 30 days after the code ends), and which account the code made and when; a sign-up refused for want of a usable code is counted by its reason only (no address, email or code), for 90 days. Named SnapAI staff see these invitation facts, and the name and email of the accounts each code made, in SnapAI's private console, which they open only through SnapAI's own Microsoft sign-in; every action there is recorded. The name and email are used to greet you during setup and to notice when you try to create a second account with a Google-verified email that already has one (you're asked to sign in instead; accounts are never joined). The About you choices are kept so setup carries on in any browser and the workspace opens only once they're complete. It never holds meeting content, documents, notes, access to your drive (each browser keeps its own connection), your IP address or your browser's details. On the website the same requests as in the extension go to the AI service, with the same content, carrying a five-minute token from this account instead of your Google or Microsoft sign-in. Signing out of Udumbara there sends any changes to your drive first, then ends that browser's session, forgets its drive connection and deletes that browser's copy of your records. You can ask for the account to be deleted from the website; it is deleted within 30 days. The extension can join the same account (you press Join, and sign in as on the website): it then keeps a session of its own on this computer, in the extension's own storage, and its requests to the AI service carry a five-minute token from the account instead of your Google or Microsoft sign-in; leaving ends that session. Joining also records the drive the extension uses with the account (the same facts as above).
Asking to join the beta (on the website)
When you ask to join the beta on Udumbara's website, the form sends your email address, whether you'd use Udumbara for business or personal work, and how you heard about it. The website passes them on as one email to SnapAI's inbox (info@snapaisolutions.com), with your address to reply to, through Resend, the email service SnapAI uses, in the United States; Resend keeps a copy of sent emails for a limited time under its own terms. The website keeps none of it, and logs only whether the email went. SnapAI uses it only to reply about the beta, and deletes it when you ask, or when it's no longer needed for that. To guard against abuse, the number of requests from one internet address is limited for a minute; that address isn't kept.
Your choices
- Captions are read only while you are in a call with captions on, and only after you've connected a drive.
- When you join a call, Udumbara turns the meeting's captions on for you the way you would: Meet's own captions button, or in Teams the call's More menu › Language and speech › Turn on live captions. If captions go off partway through the call, it turns them back on once and says so; if they go off again, it leaves them off. The captions are shown only to you; other participants aren't told. If the meeting's organiser or organisation has turned captions off, Udumbara says so and doesn't try to get round it. Switch all this off with "Turn on captions automatically (Meet and Teams)" in the side panel.
- Audio listening is off until you turn it on, and even then it starts only when you press the Udumbara icon in that call. Pause stops it (the audio streams are stopped, nothing from those minutes is kept). The first time it starts in a call, Udumbara drafts a note into the chat box for you to send, telling the others.
- Remove a document from an engagement's Knowledge tab: it and the text saved from it go to your drive's recycle bin or trash, and suggested answers stop using it straight away.
- Remove the extension to delete its data from your browser. Your drive copy stays until you delete it.
- Tell meeting participants you're using an AI assistant that reads captions — and, when you turn it on, listens to the call audio — as your organisation's policy and local law require.
Contact
SnapAI Solutions, 80 Atlantic Ave, Toronto, ON M6K 1X9, Canada. Privacy questions and requests: info@snapaisolutions.com. We will post changes to this policy here, with the date.