On-Device AI: What Stays on Your Mac, and What Doesn't
When you use a cloud-based meeting tool, here's what happens: your audio gets captured, uploaded to a server you don't control, processed by an API you can't audit, and stored in a database you have no visibility into. Some tools keep recordings for 90 days. Some keep them indefinitely. Some use your data to train their models. Most don't tell you which.
For a personal conversation, maybe that's fine. For a business conversation where you're discussing deal terms, client strategies, competitive intelligence, or personnel decisions? That's a different calculation.
The Cloud Model
Most AI meeting tools follow the same architecture:
- A bot joins your call (or you upload the recording)
- Audio is sent to a cloud transcription service (typically a third-party API)
- The transcript is processed by a language model (often another third-party API)
- Results are stored in the vendor's database
- You access your data through their web app
At minimum, your conversation passes through 3 different systems owned by 3 different companies. Each one has their own privacy policy, data retention rules, and security posture.
This isn't necessarily bad. Some companies do this responsibly. But you're trusting a chain, and a chain is only as strong as its weakest link.
Where Sussur Draws the Line
Sussur is a hybrid, and we would rather be precise about it than sell you a slogan.
Parakeet for transcription — on your Mac. Parakeet is a speech-to-text model that runs directly on your Mac's hardware. It converts audio to text without any network request. No recording file is uploaded; the audio is processed in memory and discarded. Your audio and video never leave your device.
Rapport signals — from the transcript. The rapport score is computed on your Mac from the text of the conversation: turn balance, response latency, question reciprocity, and four other linguistic signals. Sussur reads nothing from your camera, and no video frame is transmitted anywhere.
LanceDB for context retrieval — on your Mac. When Sussur surfaces a coaching tip, it searches across frameworks (negotiation tactics, rapport signals, past meeting context). LanceDB is an embedded vector database that runs locally. No cloud vector store. The search happens on your SSD.
The coaching model — a network call, with text. This is the part that leaves. To generate a suggestion, Sussur sends transcript text and the context around it to the AI provider whose API key you configured. Not audio, not video, and not a full session archive — but it is your words, and it goes over the wire. The request goes straight from your Mac to that provider: there is no server of ours in the path, and what they do with the text is governed by their terms. Run a model under Ollama instead and this step never leaves the machine either.
Storage — on your Mac. Transcripts, reports, contact personas and rapport scores are written to SQLite and LanceDB on your filesystem. Sussur never stores your transcripts in the cloud.
What "Private" Actually Means
Plenty of tools claim to be "private" or "secure." Those words are doing different jobs:
- Encrypted in transit means data is encrypted while it travels over the internet. This is the bare minimum; every serious service does it.
- Encrypted at rest means data is encrypted on the server's disk. That protects against a specific attack — someone walking off with the hardware — and does nothing to stop the company itself from reading it.
- On-device processing means the data is never transmitted in the first place. There is nothing to encrypt in transit because nothing is in transit.
The honest description of Sussur is that the heaviest and most sensitive stream — the recording of your voice and your face — is on-device, and the coaching step is not. That is a real, checkable difference from a tool that uploads the call itself. It is also not the same as claiming that no data at all leaves the machine. Any product with an AI coach in it that sells you that stronger line is either running the model on your hardware or being careless with the sentence.
The Tradeoffs
On-device AI isn't free of tradeoffs:
- Hardware requirements. Running ML models locally requires decent hardware. Sussur requires an Apple Silicon Mac (M1 or later). If you're on an older Intel Mac or Windows, you'll need to wait for broader platform support.
- Model size. Local models are typically smaller than cloud models. Parakeet is excellent for transcription but it's not the absolute largest model available. The quality is more than sufficient for real-time coaching, but it's a design choice.
- Feature velocity. Cloud-based tools can update their models and pipelines instantly. On-device updates require app updates. We ship frequently, but there's an inherent lag.
We think these tradeoffs are worth it. Your meeting conversations are some of the most sensitive data in your business. Keeping the recording on your device is the right default, even when the coaching step still needs a model.
How to Evaluate Meeting Tool Privacy
When evaluating any meeting tool, ask:
- Where is my audio processed? (On-device or cloud?)
- Is my audio stored? (If so, for how long? Can I delete it?)
- Is my data used to train models? (Most cloud tools do this unless you opt out.)
- How many third parties touch my data? (Transcription API, LLM API, hosting provider, analytics provider?)
- Can I verify the privacy claims? (Open-source components? Audit reports? Network traffic inspection?)
If the answers aren't clear, that tells you something.
Want a coach that keeps the recording on your machine? Download Sussur for Mac. Your audio and video never leave your device.