Coming from Strudel

If you write Strudel, most of your muscle memory works here: note('c3 [e3 g3] <a3 b3>'), stack(), jux(rev), .every(4, rev), euclidean (3,8) rhythms, chord symbols, scale names, setcpm(). What follows is the honest list of differences — the places where undertone deliberately behaves differently, has a smaller surface, or adds something Strudel doesn't have.

One thing to know up front: undertone is a clean-room, MIT-licensed implementation built from the published TidalCycles papers, standard music theory, and Strudel's public documentation and observable behaviour — never from Strudel's AGPL source, sample manifests, soundfont data, voicing dictionaries, or impulse responses. That's why it exists as a separate library (AGPL is a non-starter inside most shipped games), and it's also why several of the differences below are differences.

It's a library, not a live-coding environment

Undertone is something you import into a game or app. There's no REPL, no $: lines, no hush() — you hold patterns in variables, and loop() returns a handle:

const handle = pattern.loop({ bpm: 120 });
// ...later:
handle.stop(); // stops scheduling; already-scheduled voices (~0.3 s lookahead) ring out

The flip side is .play(), which has no Strudel equivalent: it schedules one cycle, once, with percussive envelopes — the game-SFX path. A consequence worth internalizing: .play() runs cycle zero only, so a <a b c> alternation in a one-shot always plays its first option. Alternation is a looping construct.

Numbers in note() are Hz, not MIDI

note(60) in Strudel is middle C. In undertone, a bare number is a frequency in hertz — note(440) is concert A, note(60) is a 60 Hz rumble. Use note names ('c4', 'f#3', 'bb2') when you mean pitches; noteToMidi/midiToFrequency are exported if you need to convert.

Relatedly, n() means exactly one thing here: scale degrees, resolved by .scale('D5:minor'). It never selects a sample variant — there is no s("bd").n(2)-style sample indexing, and no chainable .n() at all. One name per sample; .bank() handles kit-prefix lookup.

Patterned parameters: point-sampled at each onset, strings only

Every numeric control (gain, lpf, pan, envelope stages, effect sends, ...) accepts a mini-notation string, but the input is only ever a number or a string — you can't pass an arbitrary Pattern, and there are no continuous signals (sine, saw, rand, perlin don't exist).

The combination rule is simple and worth stating precisely, because it's the semantics everything else follows from:

• Structure always comes from the pattern being controlled. A four-step .gain('.5 .2 .35 .2') on a two-note pattern never creates extra events.
• The control is sampled at each event's onset. The control string divides the cycle its own way; at each event's onset instant, whichever control step covers that instant supplies the value.
• A rest (~) in the control leaves the value unset for that event — the default (or an earlier chained call's value) applies. .lpf('400 ~') gives the second event no filter, not a cutoff of 0.

Also: every control value is validated eagerly when the pattern is built — a typo'd note name, an out-of-range pan, or a malformed number in a control string throws at construction time, not mid-performance. There is no randomness anywhere in the pattern engine (no degradeBy, no sometimes, no ?); the only nondeterminism in the whole library is the noise oscillators' sample data.

.voicing() is a deterministic approximation, not voice-leading

Strudel voices chords through voicing dictionaries. Undertone's .voicing() is a stateless, pure function of the chord symbol and an optional anchor:

• The root lands within a half-octave of the anchor (default middle C; .voicing({ anchor: 'g4' }) to taste).
• Every other tone stacks directly on top by its interval, so extensions like a 9th land an octave up rather than folding down next to the root.
• Because it's a pure function, chords sharing a root and overlapping intervals (Dm then Dm7) automatically land their common tones on the same pitches — an approximation of voice-leading, with no guarantee across different roots. Dm to G7 may leap where a dictionary voicing would glide.

There's no mode, no dict, no voicing-dictionary system — { anchor } is the entire options surface. Same input, same pitches, every time, which is exactly what you want when a sound has to be identical across every player's machine.

Samples: strictly bring-your-own-assets

Strudel gives you a default sample map and soundfonts from its CDN. Undertone ships nothing and fetches nothing by default — not Strudel's CDN, not anyone's:

registerSamples({ bd: '/audio/kick.wav', sd: '/audio/snare.wav' });
const ctx = new AudioContext();
await loadSamples(ctx); // optional preload; otherwise decode happens on demand
s('bd sd').loop({ ctx, bpm: 120 });

registerSample(name, source) takes a URL, an ArrayBuffer, an AudioBuffer, or { url?, data?, buffer?, baseNote? } (baseNote being the recording's actual pitch, for pitched playback via note(...).s('mySample')). A sample that isn't decoded yet is skipped for that onset — with a once-per-name console warning — rather than throwing mid-loop. You're responsible for the rights to whatever you register.

Effects live on shared orbit buses

Reverb and delay are sends to per-orbit buses, not per-voice inserts — undertone builds a fresh node graph per onset, and a convolver per voice would be unshippable. .room(level)/.delay(level) set the send; .roomsize(), .delaytime(), .delayfeedback() configure the shared bus, last-writer-wins across every voice on that orbit. When two layers need genuinely different rooms, give them different orbits:

const pad = chord('<Dm9 Gm9>').voicing().sound('sine').sustain(0.8).room(0.5).roomsize(8).orbit(1);
const pluck = n('0 4 2 7').scale('d4:minor').delay(0.3).delaytime(0.25).orbit(2);

Unlike Strudel's orbit control, undertone's .orbit(n) takes a plain non-negative integer and is the one voice control that is not patternable — it selects a bus rather than shaping a value. The reverb's impulse response is procedurally generated from roomsize (no IR files), regenerated only when that orbit's size actually changes.

The per-voice effect roster is intentionally small: lowpass with a full ADSR filter envelope (lpf/lpenv/lpa/lpd/lps/lpr), static highpass (hpf), a 4-stage phaser, pitch slide, and the orbit sends. No distortion, bit-crush, vowel filter, chop, or the rest of the superdough rack.

.pan() runs −1..1, not 0..1

Strudel pans from 0 (left) to 1 (right) with 0.5 centre. Undertone uses the Web Audio convention: −1 left, 0 centre, 1 right, validated at build time. Porting a Strudel pan value means pan(2 * x - 1). This is also what jux does under the hood — original at pan(-1), transformed copy at pan(1).

Beyond stereo, undertone adds placement Strudel doesn't have: .channels([...]) for raw per-speaker gains up to 7.1, .surround(angleDegrees) for equal-power placement on the speaker ring, and enableMultichannel(ctx) to opt into real surround hardware (everything folds down to stereo otherwise).

The mini-notation subset

Supported, with Strudel-compatible meaning: sequences, ~ rests, [a b] subdivision, <a b> alternation, a,b parallel, a*2/a/2, a!3 replication, a@3 elongation, and a(3,8,rot?) euclidean rhythms.

Not supported: ? (degrade), | (random choice), {a b c}%n (polymeter), and the . grouping shorthand — the random ones deliberately (the engine is deterministic), the rest simply not yet. A string using them throws at parse time, so ported patterns fail loudly rather than silently drifting.

Tempo

play()/loop() take { bpm }, counting four beats per cycle — the default 120 bpm means 2-second cycles, which lands at the same speed as Strudel's default (0.5 cycles per second). setcpm(cyclesPerMinute) sets a default for calls that omit bpm (an explicit bpm always wins), and resetTempo() clears it. There's no .cps() or per-pattern tempo — tempo belongs to the play()/loop() call, not the pattern.

The shape of what's missing

Undertone's transform vocabulary is the core set: fast, slow, rev, every, euclid, jux, plus stack/seq/cat/arrange and the lower-level mini/pure/silence/timecat. If your Strudel style leans on sometimes, off, ply, chunk, add, or signal modulation, you'll be reformulating rather than porting. Synth sources are four oscillator waveforms plus three noise colours (sine, triangle, square, sawtooth, white, pink, brown) — no soundfonts, no wavetables, no ZzFX.

What you get in trade: a dependency-free, tree-shakeable MIT library with exact rational timing, eager validation, first-class one-shots, surround placement, and a fully unit-tested audio graph — built to be shipped inside something, not performed in. Start with getting started, or go straight to the recipes and the API reference.