The beginning of agentic DAW systems.

Render. Hear. Correct. Repeat.

The engine does not stop at a plausible score. It renders audio, extracts acoustic features, audits that essential chord tones (the 3rd and 7th) are audible, and automatically corrects omissions.

Text-native synthesis, 60+ years of canon

Csound is text-native synthesis built on sixty-plus years of computer music canon. Its power was never in doubt — its cost of entry was. We built a way in: describe the sound, read what got written, keep the file.

From a sentence to a p-field

Nothing here is a metaphor for what happens — this is what happens. The phrase maps to a typed edit, the edit lands on a named p-field, and the p-field is a number in the score line Csound reads. Pick a phrase; the lines it moves light up.

One parameter at a time, rendered and measured

Four compiles of the same eight bars, each moving exactly one FM p-field off its default. Same notes, same filter, same seed. Each player below holds the first twelve seconds of a render made by Csound 6.18 from the .csd beside it. The proof matrix reports spectral centroid and decay span for Default, Hollower, More Edge, and Let It Bloom.

Measured on the opening chord — four voices, 1.6 s window: centroid is the spectral centre of mass — the honest proxy for "bright." Decay span is the centroid at the attack minus the centroid late in the chord: a big span means the tone opens bright and closes dark, which is what a bell does. Note the instrument's own low-pass sits at 3 kHz, so this is a warm Rhodes by design — none of these edits are trying to be glassy in the DX sense.

The synthesis surface you can talk to

Every voice parameter below is a real p-field on a real instrument — reachable by sentence, by knob, or by API, and identical in all three paths because they write the same field.

An honest boundary: the one-word chips are mix and arrangement moves, not sound design. "Dreamier" and "punchier" push bus parameters — reverb feedback and return, delay feedback, duck depth — and layer levels, which reach the score only as each note's amp p-field. "Darker" pulls the reverb cutoff and the master lowpass down. Per-voice timbre — cutoff, drive, resonance, ratio — moves through the sentences and knobs above. We'd rather tell you where the line is than let you find it.

The conversation doubles as the lesson

A student who cannot yet write an orchestra can describe one, then read what was written for them — with their own words still attached to the lines those words moved. The file that comes out is not a toy export; it is real Csound, emitted the way a machine emits it — the standard table set and the bus plumbing spelled out whether or not a piece leans on them — and it runs anywhere Csound runs.

The agent works from the same material a student would: it can pull any voice's actual orchestra text mid-conversation — the same code you can download here. Chord harmony is walled off from it entirely: the agent names chords by symbol, never sees a voicing, and the THIRI engine resolves each symbol to pitches server-side — chord tones are computed, not guessed. Melodic lines are the exception, and we'd rather name it: solos and licks do let the agent place explicit note numbers, and the compiler honours them verbatim.