Questions, answered

No. Recording, transcription, and speaker identification all run on your machine. You can use Daisy day-to-day with no key and no account.

You only need a key if you want Daisy to write summaries, answer questions across your meetings, generate chapters, or run Coaching feedback for you. Even then, you can skip the key entirely and use the Copy all button on a transcript to paste it into your own ChatGPT or Claude.

First, you have three ways to run the AI step at all: a local model on your own hardware (nothing leaves), your own cloud key, or Daisy Cloud (managed keys, no key of your own to manage). This question is about the cloud-key route.

Anthropic (Claude) is the recommended default — it produces the strongest long-meeting summaries and structured action items. OpenAI (GPT) is a solid generalist; familiar if you already have an OpenAI account. Groq is the fastest and cheapest, with a usable free tier — great for trying things out, smaller model catalog.

You only need one. You can switch later without losing anything. Step-by-step setup for each provider.

Open Daisy and go to Settings → Providers. Pick the provider you generated a key for under AI provider, click Configure, paste the key, click Test key & list models →, pick a model, and Save.

Daisy stores the key in your local encrypted vault — the same vault that holds your voiceprints — and never sends it anywhere except the provider you picked.

Pennies. A full meeting's AI — summary, chapters, coaching, and goals — works out to roughly 4¢ for a 1-hour meeting on a balanced mix of efficient models, and up to about 30¢ if you force a top-tier model for everything. Those are approximate and vary by provider; the charge goes straight to your own key, and Daisy never proxies or marks up.

As soon as your key works, hop back to the provider's console and set a monthly spending cap (Anthropic: Settings → Limits; OpenAI: Settings → Limits; Groq: Settings → Billing) so a runaway bill is impossible. A $10–$25 monthly cap is plenty for solo use. The price of Granola or Otter — but with all of the control.

Onto your own disk, and nowhere else by default. Recordings live as files in a folder you choose (.wav chunks plus a compact stereo .opus archive). Recording, transcription, speaker identification, and search all run on your machine.

The only ways anything leaves: you press “Send to…”, you opt into a cloud model for the AI step — your own key or Daisy Cloud — in which case only the transcript text goes, over HTTPS (never the audio), or — on a lighter laptop — you choose a Deepgram key for live captions. That's it. Private by default, your choice to extend.

Daisy Cloud is an optional, managed way to run the AI step — transcript cleanup, summaries, and asking questions across your meetings — without bringing your own API key. Daisy owns and manages the provider keys, so there's nothing for you to set up or top up.

Recording, transcription, and speaker labels still happen on your machine; only the transcript text is sent — never your audio. Traffic is proxied straight through to the provider, and no conversations are stored on the way.

It's a limited public release, free for now, with no warranties. More on Daisy Cloud.

Your recordings, transcripts, summaries, and notes stay on your disk in a folder you choose. Nothing is uploaded unless you explicitly send it or pick a cloud provider.

Of course, if you keep that folder in a cloud-synced location — Dropbox, OneDrive, Syncthing — every device can share the same profile and recordings, but those files now also live on that cloud. Your call, and out of our hands.

Mostly. Recording, transcription, speaker labels, and search work with the network off, and they keep working through extended periods offline. Summaries and Q&A need a language model — run a local one (Ollama or LM Studio) and even those stay fully local; bring a cloud key or use Daisy Cloud and only the transcript text goes out when you have a connection.

In the spirit of transparency — we don't want your data — here is every time Daisy touches the network:

• It checks your license key's validity at least once every 24 hours.
• If you use Daisy Cloud, your license key is sent with every request so we can confirm eligibility.
• Update checks and device activation.

That's the whole list. Your meetings, audio, and transcripts are never part of any of it.

Daisy writes a quiet log file to your profile folder (logs/daisy.log) with stage timings and error context — enough to pin down a slow finalize or an unhappy mic without ever leaving your machine. Logs rotate daily; the last 7 days are kept and older rotations are deleted.

Nothing ships anywhere automatically. There's an Open logs folder button in Settings → Behavior, plus a Verbose debug logging toggle if you're chasing something specific. Diagnostics & logs.

The everyday proof is the design itself: Daisy is local-first, so your recordings, transcripts, and summaries sit on your own disk in open file formats (.wav, .opus, plain text) that you can read and move without us. You don't have to take our word for where your data is — you can see it.

Beyond that, an independent code and security audit is available on request. The requesting organization pays Daisy's reasonable costs of supporting the audit, plus a 10% administrative fee; the auditor is mutually agreed and works under NDA; and Daisy reserves the right to publish the resulting audit report together with the name of the requesting organization.

When you reclaim space under Settings → Recordings, Daisy removes the bulky raw WAVs but keeps a compact stereo meeting.opus archive (your mic on one channel, the system audio on the other). You can re-transcribe an old meeting later — or re-label the speakers — from that archive, best-effort.

Deleting a meeting outright from the Library wipes the Opus archive too — that one is final.

Those tools process your audio in their cloud, store your transcripts under their terms, and run them through a model you don't control — most join your call as a bot. Daisy keeps recording, transcription, and speaker identification on your machine, with no bot and no account, and lets you bring your own model or key.

See the side-by-side comparison →

Mostly yes. Recording, transcription, and speaker identification all run on your machine out of the box — the speech-to-text model is bundled, no download or key needed.

Summaries are the one piece that always needs an AI service. You can run it three ways: your own API key with a cloud provider (Anthropic, OpenAI, or Groq), Daisy Cloud (managed keys), or — if you want fully offline — a local AI server like LM Studio or Ollama, the advanced path. Or skip the provider entirely and copy the transcript into your own ChatGPT for free.

Daisy identifies speakers right on your device. Label someone once and it will recognize their voice in future meetings. Voices are encrypted, local to your machine, and never uploaded to us or anyone.

In long meetings where a person's voice drifts (they move around, switch rooms, get tired), Daisy collapses what would otherwise be two clusters back into one identity as long as both match the same vault voiceprint. You don't have to merge by hand.

Your finished transcript and speaker labels are free and produced on your machine — no per-minute bill or API key. (Live captions on a lighter laptop can optionally use a paid Deepgram key; the saved transcript is still local.)

The summary and other AI features are where a model comes in. Run a local model (LM Studio or Ollama) and it's free, on your own hardware. Or use a cloud key: a full meeting's AI — summary, chapters, coaching, goals — works out to roughly 4¢ on a balanced mix of efficient models, up to about 30¢ if you force a top-tier model for everything. Those are approximate and vary by provider; the charge goes straight to your own key — Daisy never proxies or marks up. You can also skip it and copy-paste the transcript into your own ChatGPT or Claude, free.

It depends on your machine — and it's your choice. Live captions (words appearing as people talk) are the one hardware-sensitive part, because they have to keep up in real time:

• Desktops and Apple Silicon Macs run live captions on-device — private, nothing leaves the machine.
• Lighter laptops choose: on-device captions (heavier on the machine), a bring-your-own Deepgram realtime key for a lighter footprint, or skip live captions and take the full transcript when the call ends.

That Deepgram key is the only place the cloud can touch transcription, and only for live captions on lighter laptops. The finalized transcript is always produced locally, and diarization (who said what) is always on-device and free, in every configuration.

Recordings rotate into bounded chunks on disk every few minutes, so an unexpected crash mid-meeting costs you at most the open chunk — not the whole session. Reopen Daisy and the in-progress session is recovered automatically.

If the app window goes away while a meeting is finalizing, a background process keeps the local work going. When you reopen Daisy, a corner toast catches you up: “Finalized” with a click-through to the session.

It's expected. Daisy records two independent streams — your microphone (Me) and your system audio (Them) — and your mic can pick up the other person's voice through your speakers, so the same sentence gets transcribed on both sides.

You don't need to do anything. When the call ends, Daisy reprocesses the recording, matches it against your voiceprints to label each speaker, and merges those duplicates into a single attributed turn. The saved transcript is the source of truth; the live view is a preview. Wearing headphones avoids it almost entirely. More on live transcription.

Transcription runs live, on your machine — the words appear as people speak. On most hardware it keeps up with the conversation in real time, so there's nothing to wait for when the meeting ends.

Speaker labeling runs on-device too and is fast on all three.

The AI features (summary, chapters, coaching, asking questions) are the part that takes a moment. A local model writes an hour-long meeting's summary in about 2–3 minutes (a roughly 12,000-token transcript), depending on your machine's speed; a cloud key returns it in seconds. It all runs in the background, so you can close the app while it finishes. Full speed breakdown.

Yes — end a search with a question mark and Daisy answers from across your meetings, with citations linking back to the exact moment each point came from.

Unlike keyword search, asking questions needs a language model, because Daisy has to read and reason over your transcripts to compose the answer. That can be a local model (so nothing leaves your machine) or a cloud key you bring — the same provider setting your summaries use.

Keyword search — yes, completely. It's fully local, with no AI and no network.

Asking questions needs a language model. With a cloud key that step needs internet; with a local model it works fully offline. Install LM Studio or Ollama and run a compact model — something like Gemma 3n E4B is small enough to run on a MacBook — and you can ask questions across your whole history with the network off.

Since Daisy never took your email, there's no reminder in your inbox — so this is the path back. Your meetings are still on your disk in the folder you chose; they were never on our servers. Re-download Daisy, point it at that folder, and everything is where you left it.

Pick up where you left off →

No. During the public beta the Windows builds aren't code-signed with an EV certificate yet, so Windows SmartScreen (and sometimes your browser) flags the download as “not commonly downloaded.” This is normal for any new, unsigned binary — not a sign anything is wrong with the file.

You can verify the SHA-256 of the file against the published checksum on the changelog before running it. Signed builds are on the roadmap. Full Windows install guide.