# OpenMila > OpenMila is an independent community port of Mila, the on-device > transcription app for macOS, to the Linux desktop and to Windows. It records > meetings and dictation, transcribes Hebrew and English locally with > whisper.cpp, names the speakers with pyannote diarization, writes AI > summaries through a provider you configure - including a local Ollama, which > keeps them on the machine - and exposes your transcripts to AI tools over > MCP. Free, Apache-2.0, no account, no telemetry. Copyright 2026 NX1X. Licensed under the Apache License, Version 2.0. This file summarises https://openmila.nx1xlab.dev for answer engines. Every claim here is also on the site, and traces to the repository's own README, docs/port/PARITY.md, docs/openmila/INSTALL.md or SECURITY.md. ## What it is - A port, not a new product. OpenMila copies Mila's features rather than inventing new ones. Mila was created by Uri Harduf at Island Technology, Inc. and released under the Apache License 2.0 at https://github.com/island-io/mila - **OpenMila is not affiliated with or endorsed by Island Technology, Inc.** - On macOS, the answer is to use the original Mila. OpenMila does not build for macOS and does not try to. - Repository: https://github.com/NX1X/OpenMila - Version scheme: `+port.`. It currently tracks Mila v1.9.5-beta.2; the port's own version is 1.9.5-beta.2+port.3. ## Status, stated honestly - **Released as a beta.** Downloads for Linux and Windows are at https://github.com/NX1X/OpenMila/releases and linked from https://openmila.nx1xlab.dev/download - **Linux is the development platform** (Ubuntu 26.04, GNOME on Wayland and X11, no GPU). Evidence for Linux features is direct: a test, the CI pipeline, the headless self-test, or a run of the app. - **Windows is built and tested in CI** on a `windows-2025` runner, where packaging starts the application itself before it writes a package. Hands-on testing on a Windows desktop has only just begun, so Windows features are marked as written rather than done until that catches up. - The per-feature table, with evidence, is at https://openmila.nx1xlab.dev/features and in https://github.com/NX1X/OpenMila/blob/main/docs/port/PARITY.md ## What it does - **On-device transcription** with whisper.cpp: on a Vulkan GPU where the machine has a usable one, on the CPU where it does not, which is fully supported and the common case. A software-only Vulkan device is refused because it is slower than the CPU backend. No GPU is needed, no speed-up has been measured yet, and there is no CoreML or Neural Engine path off macOS. - **Hebrew and English.** ivrit.ai large-v3 for Hebrew (3 GB), OpenAI large-v3-turbo for English and other languages (1.6 GB), routed by the recording's language. Models are downloaded on first launch and checked against a pinned SHA-256. Hebrew renders right to left per block. - **Recording**: microphone, system audio, or both mixed to one mono 16 kHz WAV. On Linux, one chosen application's audio can be captured through PipeWire. Windows has the same choice through WASAPI process loopback, built in CI and never run on a Windows machine. - **Speakers**: pyannote diarization, live and after the fact, a colour per speaker, and opt-in deletable voice profiles that recognise a person across recordings. Off by default; enabling it downloads CPU PyTorch, about 200 MB. - **Dictation**: global hotkeys, Ctrl+Alt+2 for English and Ctrl+Alt+3 for Hebrew by default. On Linux, `XGrabKey` where an X display answers and the XDG GlobalShortcuts portal on a pure Wayland session; the in-app Dictate buttons do the same job either way. - **AI summaries and action items** through a provider you configure: **Ollama (local)**, a model server on the user's own machine at http://localhost:11434 that needs no API key, the Claude, Cursor or Gemini command-line tools, or any other OpenAI-compatible endpoint. OpenMila ships no API keys and selects no provider by default. With Ollama (local) chosen, titles, summaries, action items and the Live AI rolling summary all run on the machine and make no outbound request. - **MCP server** for AI tools: `list_recordings`, `get_transcript`, `search_transcripts`, `get_live_transcript`. Off until you turn it on in Settings; turning it off stops a running session at once. It is a consent switch, not a sandbox, because the helper runs as your user. The quickstart, with what each tool returns and the registration command for every install method, is at https://openmila.nx1xlab.dev/docs#agents ## Privacy - Local transcription never sends audio anywhere. - **The AI can be local too.** Ollama (local) is a built-in provider choice, so a complete session - record, transcribe, name speakers, title, summarise, extract action items - can happen with no network at all. - **Exactly two paths send user content elsewhere, and both are off by default**: a remote transcription endpoint the user configures, which sends audio to that server only, and a cloud LLM provider the user chooses, which receives transcript text. Out of the box no endpoint is configured and no AI provider is selected. - No telemetry, no analytics, no crash reporting, no account, no paid tier. - Audio leaves the machine only through the optional remote transcription backend, and then only to the OpenAI-compatible server you configure. - Transcript text leaves the machine only through AI features you set up, to the tool or endpoint you chose. - Other outbound traffic: model downloads, the CPU PyTorch download when you first enable diarization, and the update check against the repository's GitHub releases. Updates are never installed without you. - Secrets are stored in the desktop keyring through the Secret Service on Linux (or owner-only files where no keyring answers), and with DPAPI on Windows. - Details: https://openmila.nx1xlab.dev/privacy ## Install - Linux: `openmila__amd64.deb` (a real package: `apt` installs and removes it) or `OpenMila--x86_64.AppImage` (portable; `--install` adds a desktop entry, `--uninstall` removes it). Needs GTK 4, libadwaita, libsecret, GLib, zlib and PipeWire or PulseAudio. Recommended: `pipewire-bin` for per-application capture, `wl-clipboard` or `xclip` plus `wtype` or `xdotool` for dictation that pastes itself, `ffmpeg` (or GStreamer) for `.m4a` and `ffmpeg` for importing audio that is not WAV. - Windows: `OpenMila--win64-setup.exe` installs per user with no administrator prompt and sets up the Windows App Runtime the interface needs. `OpenMila--win64.zip` is the portable form: unzip, run `WindowsAppRuntimeInstall-x64.exe` once, then `openmila.exe`. Windows 11, or Windows 10 from version 2004, x64. - Uninstalling keeps recordings, transcripts, models and settings on both systems; deleting them is a separate, explicit purge. - Verify every download: `sha256sum --ignore-missing -c SHA256SUMS`, and `gh attestation verify --repo NX1X/OpenMila` for build provenance. - Details: https://openmila.nx1xlab.dev/download and https://github.com/NX1X/OpenMila/blob/main/docs/port/INSTALL.md ## Pages - [Home](https://openmila.nx1xlab.dev/): what OpenMila is and who it is for - [Download](https://openmila.nx1xlab.dev/download): the four packages, what a machine needs, verification, uninstalling and where data lives - [Features](https://openmila.nx1xlab.dev/features): every Mila feature with its Linux and Windows status, generated from the parity table - [Docs](https://openmila.nx1xlab.dev/docs): setup, dictation, speakers, AI providers, the MCP server, troubleshooting - [Docs as Markdown](https://openmila.nx1xlab.dev/docs.md): the same documentation as one file to hand to an assistant - [Privacy](https://openmila.nx1xlab.dev/privacy): what leaves the machine, and what this website collects - [FAQ](https://openmila.nx1xlab.dev/faq) - [Changelog](https://openmila.nx1xlab.dev/changelog) - [Source code](https://github.com/NX1X/OpenMila): Apache-2.0, on GitHub - [Releases](https://github.com/NX1X/OpenMila/releases): every build, with checksums and a software bill of materials - [Feature parity table](https://github.com/NX1X/OpenMila/blob/main/docs/port/PARITY.md) - [Install, upgrade and uninstall](https://github.com/NX1X/OpenMila/blob/main/docs/port/INSTALL.md) - [Hardware and GPU notes](https://github.com/NX1X/OpenMila/blob/main/docs/port/HARDWARE.md) - [Mila, the original macOS app](https://github.com/island-io/mila) - [NX1X LAB](https://nx1xlab.dev) and the [NXTools Collection](https://nx1xlab.dev/nxtools) OpenMila is part of ## Known limits today - Linux: a meeting in a browser tab is not detected on a pure Wayland session, because no client may read another's window titles there; the Wayland global-shortcut portal is wired but an actual key press has not been witnessed by a human; no menu bar on GNOME; no blur behind the sidebar; no Neural Engine equivalent; the managed Claude install verifies a checksum because Anthropic ships the Linux binary unsigned. - Windows: hands-on testing has only just begun, so nothing there is marked done; `.m4a` depends on ffmpeg or GStreamer being on the machine, since neither ships in the package; managed Claude sign-in needs a real terminal. - Both: the Vulkan GPU backend is built, probed and wired, and has never been timed on real GPU hardware, so no speed-up figure is claimed. ## Who makes it - OpenMila is built by NX1X LAB, https://nx1xlab.dev - It is part of the NXTools Collection, https://nx1xlab.dev/nxtools ## Contact - Bugs and requests: https://github.com/NX1X/OpenMila/issues - Questions: https://github.com/NX1X/OpenMila/discussions - Security: https://github.com/NX1X/OpenMila/security/advisories/new - Anything else: https://nx1xlab.dev/contact ## Pages - https://openmila.nx1xlab.dev/ : what OpenMila is, status, the four things that matter, the non-affiliation disclaimer. - https://openmila.nx1xlab.dev/docs : the documentation. Install per system, first launch and the model download, recording, dictation, speakers, AI features, watched folders and exports, file locations, troubleshooting, and an AI-agent quickstart for the MCP server. The same page is served as Markdown at https://openmila.nx1xlab.dev/docs.md, generated from the HTML on every build, which is the copy to read or paste if you are a model. - https://openmila.nx1xlab.dev/download : packages, requirements, verification commands, file locations. - https://openmila.nx1xlab.dev/features : the full parity table against Mila, per system, with evidence. - https://openmila.nx1xlab.dev/privacy : what stays local and what does not. - https://openmila.nx1xlab.dev/faq : short answers. - https://openmila.nx1xlab.dev/changelog : releases and version numbering.