vibejam

Formely Neurosymbolic Music Generation, Vibe Musicing, Funktion.js

Coding large language models cooks every domain specific symbolic music generation model out there if we let them express music as code. Why train a bespoke neural network on MIDI files when LLMS already understands rhythm, harmony, and structure through code? This repository is a bring-your-own-coding-agent browser-based system for music composition. To use as intended, clone this repo, load it into an LLM-augmented code editor (I like cursor & claude opus 4.5), and start jamming.

Click to watch demo on YouTube

In This Repository

File	Description
livecoding.html	The browser-based music environment. Open this in a browser.
music.js	Your composition. The server hot-reloads this on save.
server.js	Optional. Enables hot-reload and sample loading.
pieces/	Example compositions to learn from or remix.

Why JavaScript

Language models understand JavaScript deeply. MIDI is a binary protocol that appears rarely in training data. This asymmetry makes JS a superior symbolic representation:

• Native fluency - LLMs write and debug JS naturally
• Abstraction creation - helper functions, chord builders, rhythm generators
• Controllability - communicate intent at any level, from "make it groovier" to "add triplets on beat 3"
• Iterative refinement - tweak code, ask for more, repeat

Representation	Problem
MIDI	Binary format, rare in training data. No abstraction layer.
MusicXML	Verbose XML. LLMs generate malformed structures.
ABC notation	Obscure, limited training examples, no composability.
Piano roll	Flat structure. Can't express "repeat this up a fifth."
Custom DSLs	Zero-shot performance is garbage.

JS wins: massive training data, Turing complete, self-documenting, composable, debuggable.

Setup

Standalone (no server): Open livecoding.html directly in a browser. Edit DEFAULT_CODE to change the music.

With server (hot-reload + samples):

1. Configure sample directory in server.js
2. node server.js
3. Open livecoding.html

The server watches music.js and hot-reloads on save.

API

Your code returns a function (t, s) => notes[] where:

• t = tick count (16th notes)
• s = time in seconds

Note properties:

Key	Description	Default
p	MIDI pitch (0-127)	60
d	duration in seconds	0.1
v	volume (0-1+)	0.5
x	pan (-1 to 1)	0
r	reverb amount	0
o	offset in ticks	0
a	attack time	0.01
c	filter cutoff	2000
w	waveform/instrument	sine
id	exclusive note id	-
start	sample start position (0-1)	0

Instruments (w parameter)

Synths: sine, sawtooth, square, triangle

Built-in drums: drums with MIDI notes: 36=kick, 38=snare, 42=hihat, 46=open hihat, 39=clap

Samples: query:index (e.g. kick:0, hat:5). Boolean search: acoustic & snare:0

MIDI output: m:channel (e.g. m:1 for channel 1)

Example

return (t, s) => {
    const n = [];
    
    // kick on 1 and 3
    if (t % 8 === 0 || t % 8 === 6) {
        n.push({ p: 36, w: 'drums', v: 0.9 });
    }
    
    // snare on 2 and 4
    if (t % 8 === 4) {
        n.push({ p: 38, w: 'drums', v: 0.8 });
    }
    
    // hi-hat 16ths
    n.push({ p: 42, w: 'drums', v: t % 2 ? 0.2 : 0.4 });
    
    // bass line using samples
    if (t % 4 === 0) {
        n.push({ w: `bass:${t % 8}`, p: 36, v: 0.7, d: 0.3 });
    }
    
    return n;
};

Inspirations

• DreamCoder - Kevin Ellis et al. Program synthesis through wake-sleep library learning.
• Barbarians at the Gate - AI-driven research for systems. Iterative generate-evaluate-refine loops with LLMs.
• Barbarians at the Gate (1993) - James Garner plays Ross Johnson. Unrelated but excellent.
• Strudel - Live coding music in the browser. Major inspiration for the UI and approach.

Other related work:

• Neuro-Symbolic Composition of Music with Talking Points Simon Colton et al.

Cite

@software{vibejam,
  author = {Jonason, Nicolas},
  date = {January 2026},
  title = {vibejam},
  url = {https://github.com/erl-j/vibejam}
}

Future work.

• Evolve better solutions. Use CLAP/Audio preference models as fitness.
• Mixing, sound design.
• Deeper daw integration.

---

Vibecoded: the human pointed, the LLM typed.