Meeting notes that never leave your machine.

Halvern records your meetings, transcribes them and writes the summary — all on your own Mac. No account, no upload, no per-minute billing.

Download for macOS

53 MB · version 0.1.0 · Apple Silicon only · macOS 14.4+

Why it exists

Every meeting tool that writes notes for you does it by sending your meetings to somebody else’s computer. For a lot of conversations — client calls, salary discussions, anything under an NDA — that is the wrong trade, and it is not a trade most people are asked to make consciously.

Halvern does the same job without that step. That constraint is the product.

What it does

Records both sides

Microphone and system audio together, so both halves of a call are captured.

Notices your meetings

Spots a call starting in Teams, Zoom or Meet, offers to record, and proposes stopping when it ends.

Transcribes locally

Picks the engine that fits your language: a fast model for 25 European languages, Whisper for the rest.

Summarises locally

A built-in model writes the notes. Bring your own OpenAI, Anthropic, Groq or Ollama endpoint if you would rather.

The Halvern workshop view: a meeting transcript beside its generated summary. The Halvern workshop view in dark mode: a meeting transcript beside its generated summary.
The transcript and its summary, side by side. Edit either.

The privacy claim, stated precisely

Claims like this are usually vague on purpose, so here is a version you can check in the source.

Telemetry The endpoint is read from option_env!() — a compile-time variable. Unset when the binary is built, it is None forever in that binary.
Analytics client Constructed only if a key is present. With no key there is no client, and every send path returns early.
Consent Defaults to false in every branch.

A downloaded build therefore sends nothing, and cannot be made to send anything without being rebuilt with credentials it does not have.

Each of those is a file you can open, in a repository you can clone — here.

Two features do send data once you configure them: an external summarisation provider, and remote transcription. Both are off until you set them up, and the privacy policy says exactly what each one sends.

What this page collects

A product that promises your meetings stay put should say what its own website does. This one has no analytics script, no cookies and no fingerprinting — nothing runs in your browser to watch you.

This page is served by GitHub Pages, and Download sends you to GitHub Releases, so GitHub’s servers hand you both the page and the file. They log what every web server logs:

IP addressWhich implies a rough location.
LanguageThe Accept-Language header your browser sends.
ReferrerWhich page sent you here, when your browser says.
Browser and OSThe User-Agent string.
TimeWhen you asked.

Those logs are GitHub’s and we never see them. What we can read is one number, from a public API anyone can call: how many times each release file has been downloaded in total. Not who, not where, not when — a count.

So there is no identifier that follows you between visits, because there is nothing set in your browser to follow you with, and nothing on our side to follow you into. GitHub’s privacy statement covers their half.

Take it

All of it is open source, under the MIT licence. Clone it, read it, change it, fork it, ship your own version, sell that version. There is no part held back and no edition you have to pay for to get the one that works.

That is not generosity, it is the point. Everything above this line asks you to believe something about where your meetings go — and the only reason you should is that you do not have to. Build it yourself from source and the promise stops being a promise.

If you write software, you also sit in meetings you might want recorded. The same afternoon can put you on both sides of that: the person who needs the notes, and the person who can go and check what the thing actually does before trusting it with a client call.

Read the source

Before you download

HardwareApple Silicon only — M1 or later. There is no Intel build.
SystemmacOS 14.4 or later, for the system-audio capture.
First runDownloads a speech model and a summarisation model, together 1.5–3 GB depending on your answers.
PermissionsMicrophone and screen recording. Screen recording is how macOS exposes system audio; nothing is captured from your screen.
StatusPre-release, and honest about it. Signed and notarised by Apple, so it opens without a warning.