referat

Private meeting capture. Useful minutes. Your choice of AI.

Download for Windows · Product site · Documentation

referat is a Windows desktop app that records system audio and microphone input, transcribes the conversation, and turns it into a structured summary with decisions and action items. It works with Teams, Zoom, Meet, or an in-person conversation—without adding a meeting bot.

You decide where each AI step runs: on your computer, on an internal company server, or through a cloud provider. Recordings, transcripts, and minutes stay local unless you explicitly configure an external endpoint.

From conversation to decisions

1. Record	2. Process	3. Use the result
Capture microphone and Windows system audio with one button.	Transcribe, optionally identify speakers, then generate structured minutes.	Review the transcript and export the result as Markdown or Word.

Why referat

• No meeting bot. System-audio capture works across meeting tools and for in-person conversations.
• Deployment is a product choice. Use local OpenAI-compatible services, internal endpoints, OpenAI, Azure OpenAI, Anthropic, or an already authenticated Codex CLI for minutes generation.
• Local-first data handling. Meeting files live under the user's Windows app-data directory. API keys are encrypted with Windows DPAPI through Electron safeStorage.
• Useful output, not just a transcript. The default flow produces a summary, decisions, and action items, with the original transcript available for review.
• Resilient processing. Long recordings are segmented, pipeline progress is persisted, and interrupted work can resume after restart.
• No telemetry. Network traffic is limited to the AI endpoints configured by the user and the optional update check.

Deployment options

flowchart LR
    A[Microphone + system audio] --> B[Local recording]
    B --> C[Transcription]
    C --> D{Optional speaker\nidentification}
    D --> E[Minutes generation]
    E --> F[Markdown / Word / clipboard]

    subgraph Choice[You choose where AI runs]
      L[Local machine]
      I[Internal server]
      P[Cloud provider]
      K[Existing Codex CLI login]
    end

    Choice -. OpenAI-compatible APIs .-> C
    Choice -. OpenAI-compatible or Anthropic .-> E
    K -. Ephemeral text-only run .-> E

Loading

Boundary	Behaviour
Audio and meeting files	Stored locally by default; uploaded only to the transcription endpoint you configure.
API credentials	Encrypted with Windows DPAPI; plaintext keys are never sent back to the renderer.
Codex	Reuses the local CLI login; runs ephemerally with shell, apps, plugins and web search disabled.
Renderer access	Sandboxed Electron renderer with context isolation and a typed preload boundary.
Provider redirects	Authentication-bearing requests reject redirects instead of forwarding credentials.
Telemetry	None.

Read the full architecture and security model.

Features

• Microphone and Windows WASAPI loopback recording
• Swedish and multilingual transcription through configurable providers
• Optional local speaker diarization
• Customizable prompt/template for meeting minutes
• Optional minutes generation through an existing authenticated Codex CLI, without an API key
• Markdown and .docx export, plus clipboard copy
• Provider connection tests in the app
• Crash recovery and per-step retry
• Automatic update support
• Light and dark themes

The current application interface is Swedish because Sweden is the initial target market. Transcription follows the spoken language. An English interface is on the roadmap.

Download

Download the Windows installer from the latest release.

The installer is not currently code-signed, so Windows SmartScreen may show a warning on first launch. The installation guide explains the limitation and the exact steps. Windows 10 and 11 are supported.

Run the AI locally

referat talks to standard /v1 APIs. A fully local setup can use speaches for Whisper transcription and Ollama for minutes generation.

# Transcription server
docker run --rm -p 8000:8000 ghcr.io/speaches-ai/speaches:latest

# Minutes model
ollama pull llama3.1
ollama serve

Configure transcription as http://localhost:8000/v1 and minutes generation as http://localhost:11434/v1. No API key is required for either local service. For model selection and troubleshooting, use the complete local AI guide.

Documentation

• Installation — installer, SmartScreen, and first-run setup
• Configuration — providers, models, endpoints, and minutes templates
• Local AI setup — private local transcription and summarization
• Speaker diarization — optional local "who said what" labels
• Architecture — process model, persistence, recovery, and security boundaries
• FAQ and roadmap

Development

git clone https://github.com/sockulags/referat.git
cd referat
npm install
npm run dev

Validation and packaging:

npm run lint
npm run typecheck
npm test
npm run build:win

Built with Electron, React 19, TypeScript, Tailwind CSS, Zustand, and Vitest.

Project status

referat is released software under active development. The core recording, transcription, minutes, export, and update flows are available today. Known limitations and planned work are kept in the public roadmap.

License

MIT © Lucas Skog