What “local” should mean
A meeting-notes app can store a copy on your laptop and still upload the audio for transcription. A genuinely local-first workflow keeps the recording, speech recognition, recap generation, and searchable library on the computer by default. Network access may still be needed to download the app, fetch a model, and check for updates. If you deliberately configure a remote inference server, that server receives the context needed for the request.
LokalBot follows that narrower, testable definition. It has no account, telemetry backend, or LokalBot-hosted AI service. Its built-in speech and language models run on Apple Silicon, while meetings, transcripts, summaries, and search indexes live in the app's local data directory.
From call to recap
- Capture two sources. Your microphone is recorded as “Me.” A macOS Core Audio process tap captures the meeting application's output as “Them.” No participant bot needs to join.
- Transcribe on-device. Choose Granite Speech, Parakeet, Qwen3-ASR, or Whisper. The audio stays on the Mac while the selected engine creates a timestamped transcript.
- Write the recap. A built-in llama.cpp model can produce the summary, decisions, and action items locally. You can instead select Ollama, Apple Intelligence, or an explicitly approved OpenAI-compatible origin.
- Search and replay. The transcript and recap are indexed in the local library. A search result can take you back to the relevant point in the recording.
Why two audio tracks matter
Recording the microphone and meeting app separately provides a useful speaker boundary before any diarization model runs. It distinguishes you from the rest of the call, keeps the two sources synchronized, and avoids the audible joins and calendar invitations created by meeting bots. Optional neural diarization can split the remote track further when several other people speak.
The honest trade-offs
- Hardware: LokalBot requires an Apple Silicon Mac running macOS 15 or later.
- Storage: selected models download before first use. Small choices start around 0.5 GB; high-quality language models can be much larger.
- Permissions: microphone and system-audio access are necessary for capture. Other features ask separately for Calendar, Accessibility, or Screen Recording access.
- Responsibility: bot-free is not consent-free. You must follow the recording laws and policies that apply to the people in your meeting.
- Model quality: local accuracy depends on language, acoustics, model choice, and available memory. Keeping more than one speech model installed lets you switch for difficult recordings.
Who benefits most
Local meeting notes are a strong fit for confidential client work, engineering discussions, research interviews, financial conversations, or anyone who simply does not want a permanent vendor account attached to every conversation. They are less compelling when a team primarily needs shared cloud workspaces, centralized administration, or mobile capture across several platforms.