What's new — Oscillate Audio

GYRE19 Sep 2026

Ten presets, a sane default speed, and a control that no longer kills half the effect

GYRE had exactly one program — which is to say none. You opened it, saw two dozen controls and had to guess what it was for. There are now ten factory settings, and each one answers that question by name: PAD DRIFT, PING PONG, HAAS DOUBLE, TREMOLO PAN, TAPE SWAY, ORBIT, BREATH, SHIVER, MONO SAFE. They appear in the host’s program menu and in the head of the device, where the name carries a half-line saying what it is for — the knobs already say what it does.
The default loop was a quarter note. At 120 BPM that is half a second from left to right and back, which reads as an effect rather than as movement. One bar — 2.0 s — is the new default.
ASYMMETRY at full travel stopped one channel dead. The Gerzon maths was right, but at −1 the left channel lost its side signal entirely: measured swing 0.000 against 1.000 on the right. On screen one curve simply stood still, which looks like a fault rather than a setting. The control now reaches 0.090 against 0.910 — still clearly lopsided, both sides still alive.
TRIPLET and DOTTED could both be on. Together they used to give the golden ratio — a nice idea that nobody ever guessed. They now switch each other off, and the processor has a defined rule for the case where automation sets both anyway.

SILT19 Sep 2026

A second hand for time, and a pitch control that means what it says

TIME — the read head now has its own speed. Until today the playhead was pinned to the recording: SILT could scatter, but it could not stretch, which is the one thing you reach for a granular processor to do. TIME 1 follows the input as before, 0.25 stretches a second into four, 0 holds the spot while recording continues, and negative values walk backwards through time. Measured on a 0.4 s note: audible for 1.82 s at TIME 0.5 and 3.64 s at 0.25 — 1.97× and 3.92×.
PITCH at zero still produced other notes. HARMONY was a probability that overrode PITCH: three grains in four took a chord degree and ignored the pitch setting entirely, which with the factory chord meant a minor third and a fifth. Measured over 256 grains, 14.8 % of them were notes nobody asked for. PITCH is now the ceiling for everything — at zero every grain sits on the root. Measured again: 100 %.
Every grain read mono. Left and right were summed before a grain was even panned, so anything wide in the input — cymbals, room, doubled parts — arrived flat, and all the width came from panning afterwards. Grains now carry the stereo image of the source, set by a new SOURCE control; channel correlation drops from 0.55 to 0.05 on unrelated material.
Grains fall in the bar. A PAN selector with five patterns: RANDOM as before, PING-PONG alternating each step, SPREAD placing low grains left and high right, SWEEP travelling across the grid, CIRCLE going round in four.
SHIMMER in the feedback path: pitched up every pass and filtered above, so it builds rather than arrives. Interval +12, +7 or +19. On a ringing 220 Hz note the spectral centre climbs 231 → 390 → 750 → 1932 Hz.
Cleaner highs. The read head interpolated linearly, which on non-octave ratios is 5.3 dB dirtier than Hermite with the worst of it at 17.9 kHz; and nothing band-limited the material before it was read upwards, so everything above 12 kHz folded back into the audible range. Both fixed: 2.5 dB less rubbish up top. Grain level scatter went from 7.7 dB to 3.3 dB, which takes out a flutter that had nothing to do with the material.
NaN in the input silenced the device permanently. Ten new filters hang off that input now, and one bad sample would have lodged in them for good. Nonsense is stopped at the door.
A key detector was built and then removed again. It estimated root and mode from the input with 36 Goertzel filters against the Krumhansl-Kessler profiles and scored 5 of 5 on synthetic cadences — but on real material it did not convince, and an automatic you cannot trust is worse than none: you never know whether the sound or the guess is at fault. Working on it produced the more useful insight anyway: SILT shifts intervals, not absolute notes, so a root note cannot affect anything at all. Only major versus minor does, through the third in the chord pool and the scale the scatter snaps to — and that is now a single switch.
Everything above costs 3.0 % of one core instead of 4.4 %.
Reverse grains get a softer window, because reversed attacks end in the steep edge rather than the decay. No signal jumps were measurable either way — zero hard steps on a plucked note, largest 5.8 % of peak — so this one is judged by ear, not by meter.
One factory preset, BROKEN — Time Smear, inherits FREEZE and runs fully wet, so it makes no sound until something has been played through it.

SCREE, SILT & CAIRN18 Sep 2026

The interfaces now speak English everywhere, not just where you look first

SCREE spoke German below the surface. The device was built in German and translated where it shows at a glance — but the line that explains a control when you hover it was never translated, and neither was the preset menu, the four preset dialogs, the unlock window, the demo notice, or the one-line description behind each of the 28 factory presets. Around fifty pieces of text. It came to light in a screenshot of SCREE running in Ableton: along the bottom edge stood “BIAS: verschiebt den Arbeitspunkt”. Everything is English now.
SILT said “KEIN SIGNAL” when nothing was playing. One line, in the place you look when something seems broken.
CAIRN listed two rooms with German dates — Paläolithikum and bis 1980er — and wrote centuries the German way, “13.” instead of “13th”.
Yesterday’s CAIRN translation broke a comparison. The room data now says “not stated” where a source gives no figure, but the interface still checked for the German wording — so 13 of the 25 rooms printed VOLUME not stated where a dash belongs. Rooms with no date recorded also showed “· NOT STATED” next to their location; now the location simply stands alone. In the field notes table, where the phrase is an honest answer rather than filler, it stays.
A check that finds this by itself. No compiler and no test run notices a German sentence in an English interface — it only shows when someone looks. sprachpruefung.py now reads every string in all five devices and flags German by its umlauts, its common words and its word endings. Run against yesterday’s files it finds 85 lines; against today’s, none. It runs with every build, as a warning that never stops the build.

SCREE & CAIRN18 Sep 2026

Both interfaces measured, and made cheaper to draw

SCREE drew its whole window thirty times a second. The timer ended in a bare repaint of the entire surface — header, frames, labels, grain, all of it — whether anything had changed or not. Measured on the bench: 75 ms per frame, which is 225 % of one CPU core for the picture alone, before a single sample is processed. Now the standing part of the surface is drawn once into an image and only the six moving areas are refreshed.
The mid/side EQ asked its parameters inside the pixel loop. Three curves × 600 points × eight bands × five lookups, each one building a string and hashing it, and the filter coefficients allocated again for every point. Roughly seventy thousand string lookups per frame. Read once, draw after: that field went from 9.3 ms to 0.9 ms.
Letter-spaced capitals measured every glyph twice, every frame. The width of a character in a given font does not change, so it is now measured once and remembered. That single table took SCREE from 21 ms to 14 ms per full frame and CAIRN from 26 ms to 14 ms — the same change helps GYRE too.
End result: a normal SCREE frame costs 10 ms instead of 46 ms. Same picture, pixel for pixel — the test renders it and compares.
CAIRN spoke German. The whole interface, including the help page and the field notes, was in German while the device is sold in English. Every string is English now, and the room table translates place names and periods at generation time, so a regenerated table stays English.
CAIRN: the close button sat on top of the room’s volume figure; the microphone line was cut off at the panel edge; the STAGE section of the help page was truncated mid-sentence. All three fixed.
CAIRN: the room picker and the field notes had 230 and 480 points of empty space below the content. The picker now shows what a room is before you enter it, and the field notes show the measured decay curve and the reverb time per octave — from the same recording the device runs on.
CAIRN’s knobs now take their light from the same place as SILT, SCREE and GYRE: one linear gradient from the upper left, a thin rim, a soft highlight. Long values such as “measured” are squeezed to fit instead of being cut off.
Two traps worth writing down. juce::Graphics::drawImage draws with the opacity of the last colour set — the knob cap had been drawn at 15 % for months, which is why it looked flat. And in C++ a hex escape eats as many digits as it can get: "\xc2\xb112 dB" is one broken escape, not ±12 dB.
The pictures on both product pages are new, rendered from these builds. The mid/side EQ shot shows the EQ actually in use — a state you can set yourself, not a mock-up.
Measured on a Mac afterwards. The DSP is three times faster there — 7.7 % of one core instead of 22 %, at four times oversampling. Drawing is closer than expected: a full frame and a normal frame cost almost the same (10.7 and 9.3 ms), where on Linux one was twice the other. CoreGraphics fills pixels almost for free, but the drawing commands cost the same — building paths, laying out text. Clipping barely helps there; to save more you have to draw less, not smaller.
The caches were being destroyed after JUCE had shut down. Both benches ended with a mutex error after their last line. The font, grain and knob-cap caches are function-local statics, destroyed at exit — by which point the font cache they reach into is gone. The same destructors run in a plugin when the library is unloaded, that is, when you take the device off a track. They are now allocated once and never freed: a few kilobytes that stay for the life of the process, and nothing can be destroyed at the wrong moment. All three devices.
Neither device has been measured inside a DAW. The numbers are drawing times from a bench, not host load.
CAIRN’s English has not been read by a native speaker.

SILT 1.0.018 Sep 2026

A granular effect that stays in time and in key

SILT is the fifth device: a granular effect for AU and VST3 on macOS, $19. It chops what is already on the track into grains and puts them back — on the host grid, 1/4 down to 1/32 with triplets, and pitched from a chord instead of a random range.
Transient detection fires a burst of grains that read exactly where the attack sits, so a kick still sounds like a kick. Fast and slow envelope, 45 ms between hits.
Five scenes — SPARSE, CLOUD, CHAOS, FROZEN, BRUTAL — and one MORPH knob that pushes density, grain length, spray, pitch range, reverse share, fold, filter, drive, feedback and the grid all at once.
FREEZE holds the six seconds already in the buffer while the input runs on.
The grain field on the interface is not decoration: every grain the engine starts writes its position, pitch, pan and length into a lock-free ring, and the display draws the last 256 of them thirty times a second. The picture on the product page is that display, after five seconds of a 120 BPM loop — 2407 grains.
Four recordings and the dry loop they were made from are on the product page. Same engine, same source, no reverb and no mastering afterwards.
The name is unchecked. No trademark search has been done for SILT, and the page does not go live before that is settled.
The JUCE licence question is unresolved and blocks selling — SILT, and every other AU/VST3 device here.
CPU load, latency and aliasing against an oversampled reference are unmeasured. Only the standalone has been run; SILT has never been loaded into a DAW.
It has been heard by one person on one system. No manual, no presets beyond the five scenes, no Windows build.

The site17 Sep 2026

Instruments in motion, and a hero that stops wobbling

The four device fronts in the hero are now touchable: point at one and it steps forward while the others recede, and a click brings it up large with its name, its one line and its price — a second click opens its page. Escape or a click beside it closes.
The cards under “What we make” live. CUE always showed its own front; HALO, CAIRN and SCREE were still pictures. Each now draws its own instrument: HALO’s tail with the decay curve and two meters, CAIRN’s room with the measured decay ring and the spectrum flower, SCREE’s spectrum with three splits and four band tiles. The first tile on each device page shows the same view. Computed from a test signal, not a screen recording — the same numbers the rest of this site is drawn from.
New headline face (Bricolage Grotesque) and a new backdrop: contour lines of the spectrum instead of the landscape, time running left, frequency up.
The hero wobbled. Two reasons, both measured: the cards floated through a CSS custom property inside their transform, which is not interpolated and forced a full style recalculation of card, image and shadow on every frame (33 ms per frame instead of 16.7); and the stage followed the pointer through a 600 ms transition that restarted on every mouse move (spikes to 116 ms). The float is now a compositor property, the stage is eased in its own loop, and the blur on the receding cards — 50 ms per frame on its own — is gone.
The hero loaded the HALO front at 3240 px to show it at 446 px. It now gets a 1400 px copy: sharp on a 2× screen, half the bytes.
The audio examples on the SCREE page point at two files that are not there yet.

HALO 2.118 Sep 2026

The preset list opens, and you can save your own

Click the preset name in the header and the whole catalogue drops over the device: twelve tools, twenty-four measured rooms, and a column for your own. Everything visible at once — no scrolling, no paging. The arrows still step to the neighbour.
Save your own presets. SAVE CURRENT asks for a name and stores the same ten values a factory preset carries. They live in Documents/oscillate-halo-presets.txt, not inside the device, so a HALO update cannot overwrite them and you can take the file to another computer. Eight of them; cmd-click one to delete it.
The header shows your own preset's name exactly the way it shows ours — if the values match, the name is there; if you move a control, it says CUSTOM. The display still cannot lie.
The PRO chip is gone. HALO is free and complete, so a button offering to unlock it made no sense. The switch behind it is now called HALO_IST_FREI and is a product decision, not a build flag that could quietly cripple the free device.
Tried in the test bench, not yet in Ableton: the drop-down, saving, and deleting are covered by twelve automated checks, but nobody has clicked them in a real Live set.

The catalogue17 Sep 2026

Prices, and the end of trial clocks

CAIRN now costs $9.99. It was free with a tip jar. Twenty-five measured rooms and a stage you can automate around your head is worth ten dollars once, and a price that is never discounted is easier to keep than a donation button.
SCREE is $19, not $29.
HALO stays free — a complete reverb for an email address, with the tip jar on its page. It is the way into this label, and it holds nothing back.
No more download demos. At nine or nineteen dollars a trial clock costs you more time than the plugin costs money. In its place: the full interface at full resolution, an audio example, and fourteen days money back through Paddle. CUE keeps its browser demo — the real device, running in this page, nothing to install.
The clocks are still in the CUE, CAIRN and SCREE builds. They come out with the next build of each.

CUE & HALO17 Sep 2026

They now open in Ableton Live 10 and 11

Both devices were saved with Max 9. Cycling '74 state plainly that Max 9 is not supported in Live 11 or earlier, which shut out everyone not on Live 12. They are now saved with Max 8, which Live 10, 11 and 12 all ship with.
Nothing had to be given up for it: the newest thing either device uses is gen~, from Max 6. The build script refuses to forget this — the reason is written where the version number is set.
Tested here in Live 12. Live 10 and 11 should now load them; if yours does not, we would like to hear about it. Live 9 is not supported.

SCREE 1.016 Sep 2026

Four bands, eight ways to break them

First build. A native AU/VST3 multiband distortion for macOS (universal): four bands behind phase-compensated Linkwitz-Riley crossovers whose sum stays flat to 0.02 dB, each band with one of eight character units and an optional second stage behind it.
Units: TUBE, TAPE, DIODE, TRAFO, FOLD, CRUSH, RECT, SINE — each with one CHARACTER knob running a pair of opposites. The SINE unit follows the Airwindows Density principle (MIT, Chris Johnson).
Level match per band before the mix: RMS before and after the unit, 250 ms, ±14 dB — measured within 0.2 dB at −6 and −40 dBFS for six of eight units.
INJECT: the distorted sum fed back into the input, limited and low-passed; the joystick runs WET across and INJECT up.
Eight-band mid/side EQ behind the damage; oversampling 2/4/8×.
Not yet heard outside the test bench, CPU load unmeasured, no manual yet. The page is live but the product is not on sale until both are done.

CAIRN 1.011 Sep 2026

Twenty-five places, zero latency

First public build. A native AU/VST3 convolution reverb for macOS (universal), built from twenty-five impulse responses of real places from the OpenAIR library (University of York, CC BY 4.0): York Minster, St Paul's, the Hamilton Mausoleum, Maes Howe, Newgrange, a reactor hall, a railway tunnel, a mine, a lime kiln, nine churches. Two of them are reconstructions and say so.
Zero latency: the convolution runs in its zero-latency partition mode, the host is told 0 samples, the measured onset is the room's own direct sound at 0.5 ms.
Impulse responses decoded from B-format to a virtual stereo pair, trimmed at onset, their noise floor replaced by a per-band synthesised tail down to −78 dB, matched in early energy. Every field note in the plugin says what was done.
DECAY 0.25× to 4× the measured room without resampling — above 1× the tail is slowed and extended from its own last half second. Maes Howe at 2×: 0.54 → 1.75 s, spectrum unchanged.
BLOOM: modulated allpass diffusion on the wet signal — density in the first moment, movement in the tail, the two things a static convolution lacks. Lossless: at 100 % the energy is −0.7 dB, T30 unchanged, crest factor of the first 50 ms 287 → 22.
FREEZE as a cloud: eight delay lines, Householder mixing, soft limiter. −41.1 dBFS after 1.5 s, −40.0 after 5 s, fluctuation 12.8 % — no echo. The input stays open under freeze.
STAGE: the room stands anywhere around you — azimuth −180° to 180° with interaural delay, head shadow and a rear cue (shelf and pinna notch, measured −3.6 dB above 3 kHz from behind); SPACE with allpass decorrelation and modulated cross-feed; ORBIT circles the room, one lap in 50 to 5 seconds; HEIGHT lifts or lowers it (the pinna notch moves 4.8 → 10 kHz, measured). Eight-band EQ with free centre frequencies, 40 Hz to 12 kHz, curve computed from the running filter coefficients. SWING limits the orbit to a pendulum around the azimuth. Rooms level-matched (unit energy: a sustained signal comes out as loud as it went in), OUTPUT ±24 dB on the wet signal only. Eight-band EQ on the wet signal, as a curve and as handles on the dial. RESONANCE at the room's measured mode.
The interface is one dial: the measured decay arc with the live tail on it, EQ handles, the spectrum flower, the reverb time in the middle, the room photo behind glass. Field note and room chooser with all twenty-five curves.
The Deck: MORPH into a second room (power-matched, two convolutions), JOURNEY through all twenty-five places (morph pendulum, the silent side takes the next room), EARLY/LATE split at 80 ms, DUCK from a sidechain bus, DENSITY on the bloom. Each one measured in the test program. Shimmer, gate and an LFO were built, heard in Live and taken out again the next day — they pulled the device away from what it is.
Interface: knobs with dial faces, glass panels, radar sweep, the room as a business card in the header (T30 · note · volume), A ⇌ B as a double name while morphing, cone and orbit trail on the stage, progress bar for the journey, window snapping to 75/100/125/150 %, a second colour scheme (Amber) remembered between sessions, staggered first-launch sequence.
Not signed or notarised with Apple yet — macOS asks once on first launch. No Windows build. CPU not measured, so no number is stated. The interface images on the product page are rendered by the plugin itself offscreen, not photographed in a DAW.

HALO 2.011 Sep 2026

The core, rebuilt after the best

Random drift instead of sine modulation on the delay lines (the way Lexicon's Random Hall and Valhalla's Random Space do it). The first attempt held each random value too long and measured worse than the sine — still values are still modes. The final version moves every 0.25–0.7 s. Ringing on a plate setting: 15.6 → 12.2 dB.
A four-stage feedforward diffuser over eight channels in front of the tank (after Signalsmith) replaces the four serial allpasses, which ring on impulses. Nothing in it feeds back.
The mixing matrix turns slowly, and it keeps turning under freeze — a held sound that breathes instead of looping. Freeze still measures −70.0 dB after ninety seconds.
Feedback allpasses capped at 0.5; delay lines 30 % longer for modal density; all 24 rooms re-fitted (under 5 % at 1 kHz). Decay accuracy 0.6 %, shortest decay at 50 m now 0.29 s.
CHARACTER, a second page of the strip: MOD, DIFFUSION, EARLY, ATTACK, BASS XO — five new Live parameters, all neutral at their defaults.
One stage for the whole device: a star field and a depth grid behind the three fields, on the same camera as the morph body. A camera flight on every preset change. DIFFUSION as a particle cloud around the body, EARLY as flashes on every hit, MOD as a shimmer on the skin.
2.0 has not been heard in Ableton yet — 1.1 was. CPU load is not measured.

HALO 1.19 Sep 2026

The metal is out of the tail

Long decays rang: with eight delay lines a thirty-metre room has about half a mode per hertz, and a diffuse room would need three to fifty times that. The energy piled up in a few frequencies and they sang. The lines now drift slowly against each other — eight rates between 0.17 and 0.52 Hz, a swing that grows with the decay time, at most 1.4 cents of pitch movement. Measured on a 7.4-second hall, the strongest 1 % of spectral lines against the median fell from 30.6 to 24.7 dB.
Moving a read position means reading between samples, and the usual linear interpolation is a lowpass — it would have cut the 8 kHz decay by 13 % at a ten-second setting. Six-point Lagrange instead, plus a seven-tap compensator whose response is exactly the inverse of the remaining loss. Tonal balance at ten seconds: 10.60 / 10.20 / 10.15 s at 125 Hz, 1 kHz, 8 kHz.
Both are switched off under freeze, so the feedback stays exactly 1.0 there. Freeze measured −69.9 dB after three seconds and the same after ninety.
New in the toolchain: the gen~ source is now translated to C, compiled and measured against the reference implementation on every build. It is the first way to catch a typo in the DSP without opening Max.

HALO 1.09 Sep 2026

EQ, space, and a body that breathes

Eight bands of EQ on the wet signal only, fixed centres from 60 Hz to 14 kHz, ±12 dB. The curve in the field is computed from the same filter coefficients the reverb runs, not drawn. It replaces the damping field; HF damp and LF tilt moved into the strip.
Where the reverb sits: azimuth, elevation and motion, as Live parameters. On headphones the far ear gets the sound later, quieter and with the highs shaded by the head; motion circles it around you. All three are neutral at zero.
The network field is a breathing wire body now instead of a cube. Its size is SIZE, decay stretches it, density smooths it, and the signal runs waves over its skin. The light on the thread outside is where the reverb sits — drag it.
The decay ring is a control, not just a display. So is the mix bar under it.
Twelve more rooms fitted from measured impulse responses: Central Hall, the Jack Lyons concert hall, the Lady Chapel at St Albans, Clifford's Tower, a lime kiln, a stairwell, a sports hall, a wood. Thirty-six presets in total.
The diagnostic text in the tail field is gone. If the patch catches a console error it says so in one line; otherwise it says nothing.
HALO 1.1 has not been heard in Ableton yet — 0.9 was. CPU load is not measured, so no number is stated anywhere.

CUE 0.5.19 Sep 2026

Three chips, fullscreen

The toolbar is down to FULL and MIRROR. Size presets, DIM, SHOW and BLACK are gone — fullscreen is always the native resolution of the screen it sits on, bright is bright, and nobody needs a button for a black screen.
The metronome is a row of LEDs now: the running beat lights up, the one is amber. Readable from the back of the room without counting.
No more hint text in the footer — nothing on the stage screen that is not for the band.
FULL: the window goes fullscreen on whichever screen it is on. Escape (or FULL again) brings it back.
The drawing follows the window size from the device itself now; the first Live test showed a black margin right and below.
The corner bracket top left ran over the logo. Corners are gone from the window, they stay on the cover.

CUE 0.59 Sep 2026

One device, a clock instead of a cut

Free and Pro are gone. There is one CUE with everything in it. Without a serial it runs as a demo: 60 seconds of transport, then the display freezes over and asks. Reload the set for another 60.
The chip top right of the device reads DEMO 0:42 while the clock runs, UNLOCK when it has run out, PRO once it is yours.
A serial removes the clock inside the device; removing the licence brings the clock back.
The window no longer holds back size, presets, THEN line, setlist rail, mirror or show view for anyone.
Seen running in Live on 9 Sep: toolbar, Blackout, Dim, clock, bar count, setlist rail, demo clock counting down.
If IBM Plex Mono is not installed, the device uses Menlo (Mac) or Consolas (Windows) instead of falling back to Helvetica. The Plex fonts ship in the download, in fonts/, with their OFL licence.
The freeze itself and the unlock after it have not been watched in Live yet — only the clock counting down towards it.

CUE 0.49 Sep 2026

Stage tools

Toolbar in the window head while the mouse is inside: BLACK (blackout), DIM (100 / 70 / 40), MIRROR, SHOW (no title bar), size presets. Mouse out, all gone.
Time to the next section and the bar you are in, top right of the window and on the device.
Setlist rail: the next sections with their length in bars, the coming one framed green.
A spotlight behind the digit; on the last bar the digit pulses with the beat.
Portrait windows (1080 × 1920) keep head, footer and type at the size of the shorter side.
Messages for MIDI mapping and message boxes: blackout 0|1, mirror 0|1, dim 0.4.
Mirror: the chip switches in Live; whether the image and the mouse flip correctly still needs a look without Blackout on.

CUE 0.38 Sep 2026

Free and Pro

Free and Pro in one device. Free keeps the count on the device and opens the window at 1280 × 720; Pro unlocks any size, presets, the THEN line and the clean show view.
Unlock inside the device: FREE / PRO chip top right, email plus serial, checked offline, saved next to the device as CUE-Pro.license.
Serial is bound to your email and works on every computer you use.
The hint on the device no longer promises “pull it large” to Free users.
The unlock dialog has not yet been exercised on a fresh Live install. If it refuses a valid serial, write to [email protected] with the Max console output.

CUE 0.27 Sep 2026

The window is a window

The display follows the window. Drag it onto a second screen and pull the corner — the digits grow with it. HD / FHD / UHD / portrait presets on hover.
Metronome as a bar row: the current beat stands tall and bright, the one gets a green stripe. Counts correctly in 6/8 and 7/8, not only in quarters.
Bar tick marks in the progress band; corner brackets matching the cover art.
Max helper objects were visible at the bottom of the stage window. Moved out of view.
The seven-segment “1” read as “8” from a distance. It is now centred and has no ghost segments.
The THEN line showed a blank when there was no section after next.

CUE 0.16 Sep 2026

First build

Counts the bars until the next Arrangement locator, reads locator names, tempo and time signature straight from Live. No dummy clips, no setup.
Audio effect, so it lives on the Main track and passes the signal through untouched.
Double-click a tick to jump there, quantised. A single click does nothing on purpose.
Time-signature changes mid-set: Live does not document whether the API follows them. Untested. Tell us if your count drifts.

Every change, dated.