Skip to content
PascalNunPublic

About

Local-first AI desktop app for turning meeting recordings into editable protocols on your own computer.

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

LocaLog

Local-first AI for turning meeting recordings into protocols.

LocaLog is a desktop application for people who need a written record after a meeting, but cannot send the conversation to a cloud AI service.

You import an audio or video recording, review the local transcript, and use a local language model to prepare a protocol draft. You can edit the draft before exporting it. Projects and meetings keep the recording, transcript, protocol, and exports together.

The protocol is the point of the product. Transcription is the reviewable source that makes a reliable protocol possible.

LocaLog light start screen LocaLog dark start screen

The current shell uses synthetic project data. These screenshots show the visual direction in light and dark mode, not a finished release.

Download

The first alpha is on the releases page, for macOS, Windows and Linux. It is unsigned, so each system will warn you before opening it; the release notes say what to click and why. Read the untested list there before you rely on the output.

Why LocaLog exists

Meeting recordings can contain personal information, internal decisions, client details, and material that should not leave an organisation's controlled environment. LocaLog is designed around that:

  • meeting content stays on the device through the core workflow; an advanced setting can point protocol generation at a server you name, and it says plainly that meeting content will be sent there and requires an explicit acknowledgement before anything is;
  • no LocaLog account, cloud workspace, telemetry, or hosted AI service is required;
  • generated text is visible and editable before it becomes an exported record;
  • projects and meetings provide context for every recording and document;
  • the interface is designed as a writing and review tool, not a chatbot or model dashboard.

The project is open source under GPL-3.0-or-later. It builds and packages on macOS, Windows and Linux. macOS remains the main validation platform. Linux development and synthetic desktop-audio capture have also been checked on Omarchy; see Developing on Linux for the setup and remaining limits. Operating-system-specific work stays at the edges of the application.

The workflow

Project → Meeting → Import → Transcribe → Review → Generate → Edit → Export

Importing a recording is the path that works. Recording a meeting inside the application uses AVAudioEngine on macOS, WASAPI loopback on Windows and the desktop sound server's monitor source on Linux (PulseAudio or PipeWire-Pulse). The Linux recorder has been checked with separate synthetic virtual sources, including track separation, retained audio, levels and stopping. Physical-device capture and a real meeting still need validation on Linux.

Current state

LocaLog is an early working prototype. It is useful for testing the workflow and architecture, but it is not ready for production use.

The repository currently contains:

  • a Tauri desktop shell with Svelte and TypeScript on the frontend and Rust at the core;
  • a warm light/dark visual system using locally bundled Barlow typography;
  • project and meeting storage in SQLite with versioned transcript and protocol artifacts;
  • durable import, transcription, and generation jobs with cancellation, retry, and restart recovery;
  • local media probing and normalisation through supervised FFmpeg processes;
  • a whisper.cpp boundary for local transcription, with consent-gated verified model downloads;
  • a bundled llama.cpp server for writing protocols, and an Ollama provider for anyone already running one;
  • transcript review, Markdown editing, autosave, revision history, and Markdown, plain-text, PDF and Word export;
  • synthetic fixtures, evaluation harnesses, and isolated architecture spikes.

Seven sidecars ship inside the application: whisper.cpp, FFmpeg and ffprobe, the llama.cpp server, speaker diarisation and embedding, and the recorder. A locally supplied whisper.cpp executable and a user-managed Ollama server are both still accepted. Signing and notarisation, M1/8 GB performance validation, accessibility auditing, and backup/restore are open.

The main unresolved product question is protocol quality: whether the generated document is complete, factually supported, and useful after light editing.

Documentation

Start with the documentation guide. It explains which document to read for a particular question and distinguishes product goals, accepted decisions, current implementation status, and experimental evidence.

  • Product — what LocaLog is for, who it serves, and what belongs in the first version
  • Experience and UX — how the application should behave and feel
  • Visual direction — typography, colour, spacing, and visual character
  • MVP — the first complete workflow and its boundaries
  • Architecture — how the application stores data and runs local work
  • Decisions — accepted choices, open questions, and the evidence behind them
  • Current plan — what is true now and what should happen next
  • Roadmap — later possibilities, not promises for the current release

The spikes/ folders contain isolated studies. They are evidence and test oracles, not production modules. The local-only eval/ folder may contain real evaluation material and must never be committed or shared.

Running the current build

The browser shell is useful for visual and workflow development. It uses synthetic data. The native shell exercises the real storage and process boundaries when the required local runtimes are available.

Prerequisites:

  • Node.js 22.12 or newer
  • npm
  • Rust 1.97.1 through rustup
npm install
npm run dev

To run the Tauri shell:

npm run tauri dev

For a packaged build, which compiles every sidecar first and then bundles them:

npm run tauri:build

Expect this to take a while the first time; FFmpeg and llama.cpp are built from source. Model files are never bundled. They download when the feature that needs them is first used, with the size shown and, where a licence requires it, the terms shown before anything is fetched.

The complete checks are documented in CONTRIBUTING.md.

For Linux prerequisites, development without AI models, and packaging requirements, see Developing on Linux.

Contributing

The project is public while its foundations are still being proven. Before changing behaviour, read the product and architecture documents and check the current plan. Keep changes small, explain important trade-offs, and do not add real meeting material, credentials, model files, or runtime binaries to the repository.

Licence

LocaLog is free software licensed under the GNU General Public License v3.0 or later. You may use, study, modify, and redistribute it, including commercially, under the licence terms. Third-party dependencies, fonts, runtimes, models, and other assets retain their own licences.

Copyright © 2026 Pascal Nünninghoff.

About

Local-first AI desktop app for turning meeting recordings into editable protocols on your own computer.

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages