OpenMila

Linux and Windows

Your meetings, transcribed on your own machine

OpenMila brings Mila, the local transcription app for macOS, to the Linux desktop and to Windows. Same features, same models, same on-device privacy: recording, dictation, Hebrew and English transcription with whisper.cpp, speaker diarization, AI summaries and the MCP server for your AI tools. Nothing leaves your machine unless you point it at a server of your own, and with Ollama running locally even the summaries stay put.

Apache-2.0, free, no account. The first release has not been published yet: the download page lists the files that will be there and the commands to verify them.

Where the port actually stands

This is stated plainly because the difference matters when you decide whether to install it.

  • Linux development platform

    Linux is the platform OpenMila is developed on: Ubuntu 26.04 on GNOME, Wayland and X11, with no GPU. Evidence for the Linux column of the parity table is direct, from tests, the CI pipeline, the headless self-test, or a run of the app.

  • Windows built in CI only

    Windows is built and tested on a windows-2025 runner in CI. The core, the CLI, the MCP helper, the port's tests, the WinUI app and the portable zip all pass there. Nobody has yet run any of it on a Windows desktop, so no Windows feature is marked as done.

  • Release not yet published

    The repository describes itself as early development, not yet released. OpenMila tracks Mila v1.9.5-beta.2; the port's own version is 1.9.5-beta.2+port.3. Watch the releases page for the first Linux beta.

The full list, feature by feature, with what is missing or partial on each system and why, is on the features and parity page.

The whole thing can run offline, the AI included

Transcription was always local. The part people assume is not is the AI: the titles, the summaries, the action items, and the Live AI rolling summary while a meeting is still going. Those go to whichever provider you pick, and one of the built-in choices is Ollama (local) - a model server on your own machine at http://localhost:11434. Choose it and there is no key to enter, no account to make, and no request that leaves the computer.

So a complete OpenMila can be: record, transcribe, name the speakers, title it, summarise it and pull the action items out of it, with the network cable out. The only things that ever need a connection are the one-time model downloads and the update check, and the update check can be ignored.

Two paths do leave the machine, and both are off by default: a remote transcription endpoint you configure, which uploads audio to that server and to nowhere else, and a cloud LLM provider you choose, which receives the transcript text. Until you set one of them up there is no provider selected at all. OpenMila ships no API keys, no default endpoint and no provider of its own. Privacy has the full list of what can leave.

One caveat, stated rather than buried: the provider list is done on Linux and written on Windows in the parity table, and the run behind the Linux status used the Claude CLI. Nobody has yet driven OpenMila against a local Ollama end to end.

The four things that matter

On-device transcription

whisper.cpp runs on your machine: on a Vulkan GPU where there is a usable one, and on the CPU where there is not, which is the common case and fully supported. Models are downloaded once, on first launch, with a progress bar and a pinned SHA-256 check, and then the audio stays on your disk. There is an optional remote backend for any OpenAI-compatible endpoint, and it only ever talks to the server you configure. Neural VAD (Silero) and hallucination reduction come along unchanged from the upstream engine.

Hebrew and English

Two models, routed by the recording's language: ivrit.ai large-v3 for Hebrew and OpenAI large-v3-turbo for English and other languages. Hebrew text is rendered right to left per section, with the direction decided per block, and a dictation hotkey of its own.

Speakers and AI summaries

Speaker diarization with pyannote names who said what, live during a recording and afterwards, with a colour per speaker and an opt-in, deletable voice profile that recognises the same person across recordings. Summaries and action items come from a provider you choose: Ollama on your own machine, the Claude, Cursor or Gemini command-line tools, or any other OpenAI-compatible endpoint. Diarization is off by default, as it is in Mila.

MCP for your AI tools

The same MCP server Mila has, built from the same source, lets a Claude session search and read your transcripts, including the live transcript of a meeting in progress. It is off until you turn it on in Settings, and turning it off stops a running session at once. It is a consent switch, not a sandbox: the helper runs as your user and reads the same files you can.

Who this is for

People who take meetings on a Linux desktop or on Windows and want the transcript, the speaker names and the summary without handing the audio to somebody else's service. It suits Hebrew and English speakers in particular, because those are the two languages the bundled models are chosen for.

On a Mac, use the original: island-io/mila. OpenMila does not build for macOS and does not try to.

Not affiliated with Island

OpenMila is an independent community port of Mila for Linux and Windows. It is not affiliated with or endorsed by Island Technology, Inc.

Mila was created by Uri Harduf at Island and released under the Apache License 2.0 at github.com/island-io/mila. OpenMila reuses Mila's transcription engine, data model and application logic unchanged wherever the platform allows, and re-implements only what is tied to macOS. The upstream files this port modifies are listed in CHANGES.md.