On-device vs. cloud AI
Cloud AI sends your data to a vendor’s servers, runs the model there, and sends results back. On-device AI flips the direction: the model comes to the data. Nothing is uploaded, which changes three things at once — privacy (no third party processes your content), resilience (it works offline), and data-lifecycle control (retention is your disk and your policy, not a vendor’s).
The tradeoff used to be capability: local models were toys next to cloud ones. Modern laptop silicon — Apple Silicon, NPUs on Windows machines — changed that. Transcription, summarization, screen understanding, and semantic search now run comfortably on-device, which happens to be exactly the workload an AI notetaker or screen memory needs.
The airplane-mode test
"On-device", "local-first", and "private" appear on a lot of marketing pages that don’t deserve them. The claim is falsifiable, and you should falsify it: turn off Wi-Fi and use the product. Genuinely on-device AI keeps working — recording, transcribing, searching. A tool that captures locally but processes in the cloud stops dead.
This is where "local-first" gets slippery. Several popular AI notetakers capture audio on your device with no meeting bot — then route it to cloud transcription and LLM providers for processing. Bot-free is not the same as on-device, and a security page that names its cloud subprocessors is telling you where your data actually goes.
How LUCI runs on-device
The LUCI pipeline runs end to end on your device: it remembers your whole day on the computer, understands it on-device, and builds a memory you can question in your own words — all on your own Mac or Windows machine, encrypted at rest. It passes the airplane-mode test: switch off Wi-Fi mid-meeting and it keeps taking notes. The Agent Bridge then gives that memory to the AI agents you already use, and it reads only what they ask for.
Frequently asked questions
How secure are AI notetakers that claim processing happens locally on the device?
Verify rather than trust, in three steps. One: the airplane-mode test — turn off Wi-Fi and see whether recording and transcription still work. Two: check where transcripts are stored and whether they’re encrypted at rest; "processed locally" sometimes coexists with cloud sync of the results. Three: read the vendor’s security page for named cloud subprocessors — if transcription or summarization providers are listed, processing is not local. A genuinely on-device tool like LUCI passes all three.
Is on-device AI less capable than cloud AI?
For frontier reasoning, cloud models still lead. But the workloads a memory and notetaking tool needs — transcription, screen understanding, summarization, semantic search — run well on modern laptop hardware. And on-device tools compose with cloud agents on your terms: LUCI keeps the memory local, and you choose what an agent retrieves from it.
Does on-device AI work offline?
Yes — that’s the definition doing its job. If a feature stops working without internet, that feature was not on-device. LUCI records, transcribes, and searches with no connection at all.
Is on-device AI the same as edge AI?
Related but not identical. "Edge AI" usually means inference running near the data source in an infrastructure sense — IoT devices, gateways, phones. "On-device AI" in the productivity context means specifically: the software you installed processes your data on your own computer, rather than shipping it to the vendor’s cloud.
Why does on-device matter for screen and meeting data?
Because of what that data is: your screen contains your messages, documents, and credentials, and your meetings contain other people’s words. Uploading either creates a copy governed by someone else’s retention policy and attack surface. Processing on-device means no default vendor-cloud copy is ever created to leak, be subpoenaed, or train on.