Private beta · desktop & browser
It listens,
and shows you
what to play next.
Harmony Lab hears music — from an audio file, your microphone, or the instrument plugged into your interface — and shows the key, the chords, and the notes that will work over them, as it plays.
Nothing is uploaded. All of it runs on your own machine.
Access is granted by email address. Yours has to be on the allow-list before the app will let you in — so the link above will turn you away until it is.
The gap
Most tools stop at naming the chord.
There is a long stretch between playing some chords and hearing where a song wants to go. It lasts for years, for most people, and the reason is that the feedback is slow: you play something, it sounds wrong, and nothing tells you why — or what to reach for instead — while your hands are still on the instrument.
Chord detectors label what is playing. That is the easy half, and it is close to free now. None of them answer the question you are actually asking: given what I just heard, what can I play, and how will it feel?
Detection without direction is a dictionary with no sentences.
The idea
Safe, Color, Tension.
Every moment the app hears resolves into three answers rather than one label. Safe is the ground that fits. Color is the one note that changes the flavour. Tension is what pushes, and where to resolve it. They land on a fretboard, a keyboard, or a stave — whichever instrument you are actually holding.
A minor pentatonic
- A
- C
- D
- E
- G
Everything here fits the chord. You cannot be wrong on this ground.
A Dorian
- A
- B
- C
- D
- E
- F#
- G
One note is doing the work: the natural 6, F#. That is the whole flavour.
A blues, plus approach tones
- A
- C
- D
- D#
- E
- G
The b5, D# — leaned on, then let go. Shown only when the app is confident of the chord.
The unusual part
It runs entirely
on your device.
This is not a hedge or a roadmap item. The analysis — resampling, the transform, chord and key inference, the pathways — happens in workers on your own machine, in the app you launched. There is no server in the loop. Your audio is never uploaded, because there is nowhere for it to go.
- No upload Audio from a file, your mic, or your system output is analysed in place and stays there.
- No streaming APIs Nothing is looked up against Spotify, YouTube or a chord catalogue. The app works out the harmony itself, from sound.
- No account to analyse The invite gates who can open the app. The analysis itself does not phone anywhere.
- Your own AI, locally The app never calls a language model to do the listening. It exposes the song over a loopback MCP server so your Claude can read and shape it — your model, your machine.
The one exception, and it is off by default
Importing a photo or PDF of a chord chart can use Claude to read it. That path
only exists once you paste your own ANTHROPIC_API_KEY into Settings. With
no key, the feature falls back to a local staging flow and nothing leaves the device.
It is the only route out, and you have to open it yourself.
What's in it
A workspace, not a readout.
-
Play
The live surface: key and chords as they happen, pathways under your fingers, loops you can keep as you find them.
-
Instrument views
Guitar, bass, ukulele and keyboard — the same pathway, drawn where your hands are. Plus notation and a piano roll.
-
Looper
Record, overdub and edit a loop, MIDI or audio, without leaving the song you are playing over.
-
Song
Lyrics and a lead sheet for what you have built, with a Co-writer alongside them.
-
Explore
Harmonic space you can walk around: circle of fifths, a chord tree, and a chord lab for trying qualities against each other.
-
Learn
Guided lessons that ask you to play, rather than to read.
-
Library
Start from a template or a public-domain starter song, or reopen a chart you saved.
-
ORC-1
A chord synth for building and voicing chords directly — useful when you are writing rather than transcribing.
What you can point it at
Three ways in.
-
01
An audio file
Load a track and scrub it. Deterministic, repeatable, and the easiest way to see whether it hears what you hear.
-
02
Your microphone or interface
Play a guitar, a keyboard, or sing. The analysis runs on what you are doing right now, not on a recording of it.
-
03
Whatever your computer is playing
Capture the system output or a browser tab and improvise over it. Only the sound is analysed; the picture is thrown away.
There is a desktop app for macOS and Windows, and a browser build you can open on a phone. Feature coverage differs between the two — the desktop build is where everything works.
Straight about it
This is an early beta.
It is being built in the open by one person and an unusual amount of automation, and it is not finished. Rather than find these out yourself:
- The pipeline is tuned for relatively clear harmony. Dense or heavily polyphonic material can give unstable or misleading chords.
- Noise and low-quality audio lower confidence and stability.
- Key is estimated over a long window. Modulations are not explicitly modelled, so a key change may lag or be missed.
- The desktop installers are not code-signed yet, so macOS and Windows will both warn you on first launch.
- System-audio capture on Windows uses a newer path than the macOS one and has had less real-world testing.
- It does not try to beat dedicated transcription tools on raw accuracy. Direction is the product; detection is the input.
Telling us where it got a chord wrong is genuinely the most useful thing a tester can do — that is what this round is for.
Want in?
The beta is invite-only. Email me the address you want on the allow-list and roughly what you play, and I will add you and send back the install notes. There is no form here and nothing is collected — it is a mail link, and it goes to a person.