Mityu records meetings and on-site conversations, transcribes them on your device, and turns them into structured, source-linked summaries and action items — with no cloud and no server required. Built for teams that need meeting intelligence without giving up privacy, compliance, or control.
A bluedev product.
- Introduction
- Why Mityu?
- Features
- Installation
- System Architecture
- Roadmap
- For Developers
- Contributing
- License & Acknowledgments
Mityu is a Tauri 2 desktop app (Rust core + Next.js UI) that captures your meetings, transcribes them locally in real time, and generates summaries — without sending audio or transcripts to anyone else's servers. The capture → transcript → summary → store path runs entirely on your machine and keeps working with no network and no server. An optional sync/collaboration server can be added later for teams, but it is strictly additive: turn it off and the desktop app keeps working on your local data.
Mityu is designed for professionals and enterprises who must keep control of sensitive conversations. v1.0.4 has not yet completed its human-reviewed target-environment benchmark or pilot, so noisy-field, jargon, diarization, and accuracy claims are intentionally withheld.
- Privacy-first. Capture, transcription, and (by default) summarization run on your device. No cloud, no leaks.
- Use any model. Prefer a local open-source model? Great. Want to plug in an external API? Also fine. Bring your own key (BYOK) — no lock-in.
- Cost-smart. Avoid pay-per-minute bills by running models locally, or pay only for the calls you choose.
- Any meeting app. Google Meet, Zoom, Teams — or a face-to-face conversation. Mityu captures system audio, so it records regardless of platform (it is not a bot that joins your call). System-audio capture is supported on macOS (ScreenCaptureKit / Core Audio tap) and Windows (WASAPI loopback); Linux is experimental — microphone capture works, system-audio capture does not yet.
- Human-in-the-loop. AI summaries and action items are drafts bound to their source transcript segment until a human approves them — for trust, dispute evidence, and EU AI Act transparency.
- Local transcription with Whisper (
large-v3) or NVIDIA Parakeet — no cloud required. - Real-time transcript of the meeting as it happens.
- AI summaries, BYOK: choose Ollama (local), Claude, Groq, OpenRouter, or any OpenAI-compatible endpoint. API keys are stored in the OS keychain — never in plaintext.
- Import & enhance: import existing audio to generate a transcript, or re-transcribe a recording with a different model or language — all processed locally.
- Professional audio mixing: capture microphone and system audio simultaneously with ducking and clipping prevention.
- GPU acceleration: Apple Silicon (Metal) + CoreML on macOS; NVIDIA (CUDA) and AMD/Intel (Vulkan) on Windows/Linux — enabled at build time, no configuration needed.
- Multi-platform: macOS and Windows (Linux builds from source).
- Local-first storage: recordings and transcripts stay on your machine. In v1.0.4, raw meeting audio remains local until you delete the meeting.
Screenshots reflect the app UI and are being refreshed for the Mityu brand.
- Download the latest
x64-setup.exefrom Releases. - Run the installer.
- Download the
.dmgfrom Releases. - Open it and drag Mityu to your Applications folder.
Experimental — microphone only. Recording, transcription and summarization work, but system-audio capture does not: it needs a PulseAudio/PipeWire backend that is not built yet (see ADR-0022). Mityu will record the microphone and log a clear message instead of silently capturing nothing. Use it for in-person conversations; for online meetings prefer macOS or Windows.
Build from source:
git clone https://github.com/aydogandagidir/mityu
cd mityu/frontend
pnpm install
./build-gpu.shSee Building on Linux and the general build guide.
If no packaged release is published yet, build from source with the guides above.
Mityu is a single, self-contained application built with Tauri: a Rust core handles capture, transcription, summarization, and local storage; a Next.js frontend provides the UI. There is no required server — an optional, authenticated multi-tenant server is a later, additive phase.
For details, see the architecture documentation.
Mityu is developed local-first, then server-optional, with go/no-go gates (see docs/ROADMAP.md):
- Phase 1 — Enterprise local-first MVP: encrypted local store, source-linked HITL summaries, action-item extraction, search, export (PDF/DOCX/Markdown), consent & transparency.
- Phase 2 — Optional self-host server: authenticated, multi-tenant (OIDC + RBAC + Postgres RLS + audit); shared workspaces; the app still works with the server off.
- Phase 3 — Managed multi-tenant SaaS.
- Coming soon — on-device AI agents: a library of local agents that draft follow-ups and track action items — draft-only, human-approved, no autonomous actions (see EPIC F).
You'll need Rust and Node.js. For detailed build instructions, see the Building from Source guide. Repository conventions, architecture, security/privacy, and the decision log live in docs/, with contributor rules in CONTRIBUTING.md.
Contributions are welcome. Please open an issue or a pull request and follow the project structure and guidelines in CONTRIBUTING.md.
MIT License — see LICENSE.md.
Mityu is built on the open-source Meetily by Zackriya Solutions (MIT). The MIT copyright notice is preserved in LICENSE.md. Mityu is a separate product by bluedev and is not affiliated with, nor endorsed by, Meetily or Zackriya Solutions; it does not use the Meetily name or branding.
Additional thanks:
- whisper.cpp and transcribe-rs for on-device transcription.
- Screenpipe for audio-capture code we build on.
- NVIDIA for the Parakeet model, and istupakov for its ONNX conversion.
- The upstream Meetily contributors whose work (including import & enhance) Mityu inherits.