Running AI Locally: What NotoNote Keeps on Your Machine

Most meeting assistants work like this: a bot joins your call, records it, and sends the audio to someone else's cloud. NotoNote, the meeting assistant we build at FluidMind, works the other way around. It captures, transcribes, and summarizes on your own machine, and the cloud is there for sync, not for listening in.

The pipeline is local by construction. The desktop app is Tauri with a Rust backend: audio capture, voice activity detection, and speaker diarization on device; transcription through Whisper.cpp on device; notes and search in a local SQLite store with local vector search. Nothing in that chain needs a server.

Summaries and questions run on a local model too. NotoNote started with the usual approach, a system-installed Ollama, and it was a support nightmare: admin password prompts, conflicts with Homebrew installs, six releases fighting installer bugs, 1,120 lines of installer code. So we bundled the runtime instead: the app starts its own Ollama on a private port, downloads models into a directory it controls, no admin, no conflicts. And the next step is already in the codebase: FluidMind LLM, an embedded inference engine in Rust with Metal GPU acceleration that runs any GGUF model in-process, with a llama.cpp sidecar only as fallback. The direction is clear: the runtime becomes part of the app, not a dependency you install.

Local mode is not a setting you hope works. It activates when you are offline, when the backend is unreachable, or when you mark a meeting private, and a private meeting never leaves the device. There is a privacy toggle per meeting. Sync across devices is the default for ordinary meetings; privacy is one switch away.

The rest is the product around that core: no meeting bots and no awkward recording announcements, live transcription with speaker identification in any language, even mixed in one meeting, custom summary templates for decisions, next steps, and open questions, chat with a meeting with answers that jump back to the transcript, and action items tracked across meetings with a morning briefing. It works in English, Korean, and Japanese, with dedicated engines for the latter two.

An honest footnote on platforms: local mode is a Mac story today, Apple Silicon only. The Windows edition is cloud-only for now, and Linux is built but ships without an auto-update channel. Local AI is not a slogan here; it is a stack, and stacks have edges. The bet is simple: your meetings are yours, so the machine that hears them should be yours too. NotoNote is free to use at notonote.ai.