Muan

Every meeting tool on the market listens, then summarises. A summary restates the speaker’s own frame in fewer words.

Muan tests it instead. The transcript scrolls on the left. On the right, an analysis panel appends a pass every couple of minutes: what was claimed, whether it holds, what contradicts it, and what is still unverified.

Muan is a Kuki word for trust. The name is the promise: you can trust what you were told, because the app checked it while you listened.

What it does that a summariser cannot

• Names the claim, rather than compressing the paragraph it sat in.
• Checks it against a second source, and says which one.
• Says what would falsify its own finding, on every pass, without exception.
• Records being wrong as a new row, never as an edit. A finding that is later overturned stays on the record with the finding that overturned it.

Three rules the database enforces, not the prompt

These are constraints in the schema. An analysis pass that breaks one cannot be written to disk at all, which is a stronger guarantee than an instruction a model may ignore.

• A pass cannot exist without a disconfirming read. The column is NOT NULL and the check refuses an empty one.
• A finding cannot be marked supported or contradicted without a source. No source, no verdict.
• A number the machine heard, confirmed only by the machine that heard it, is forced to unverified. Transcription once misheard "fifteen" as "fifty" and the analysis passed it on. Now a finding citing a number-bearing segment needs a document or a URL behind it, or it does not get to claim anything.

Where your audio goes, stated exactly

Transcription runs on your machine. The model is faster-whisper, it runs on your CPU, and the audio never leaves the computer.

The analysis is different, and Muan will not pretend otherwise. Each pass sends transcript TEXT to the analysis provider. The app says so on the screen where you would have assumed otherwise, and it will never claim that your data never leaves your machine, because half of that sentence would be false.

You hold the key and the budget

• Your own API key, entered in Settings. It is held by the desktop shell and never enters the web view.
• A spend ceiling per session, set by you, three dollars by default.
• The guard fails closed, and checks the ceiling before the call rather than after it.
• When the budget runs out, transcription keeps going and analysis pauses. Never the other way round.

Analysis on a two-hour session costs roughly $2.75 at the default settings. That is an estimate from a measured run, not a promise, and it is the reason the ceiling exists.

How it is built

• A Windows desktop app, not a web page and not a browser extension.
• One local SQLite file per session. No account, no server, no sync.
• Capture what the computer plays, a microphone, both, a file, or a link.
• Export the transcript and the findings when you are done.

What it will not do

Muan never summarises. If a pass produces only a restatement, that pass has failed by its own definition.

Muan ships no downloader. A direct audio link is fetched on your confirmation. A platform page needs a tool you have already installed, and Muan names it rather than quietly doing it for you.

Muan does not read your hardware. It reports what it has actually measured on your machine and nothing else, because printing a plausible number would be worse than printing none.

Not on sale yet

This page is here so you can see what is coming. There is no price and no download, and nothing on this page can be bought today. When it ships you will be able to buy it here.