v2.3.0

The Remote Script is unchanged (2.1.0): no reinstall or Live restart needed.
The new build_tracks options were checked in Live 12.4.6 (2026-10-19);
audio_to_midi's placement through a clip's warp only against the fake Live.

Added

• audio_to_midi takes a preset: melodic (the default: onset 0.6,
  frame 0.3, 35 ms, melodia only below MIDI 52), chords (melodia at every
  pitch) or legacy (what it did before: 0.5, 0.3, 127.7 ms, melodia on),
  and melodia_trick, which it never passed (it was always on); given, it
  holds at every pitch. On 63 synthetic renders (plucks, 32nd runs, leads,
  pads; with reverb, an 808 and a bleed loop), onset F1 went from 0.583
  (legacy) to 0.687 with melodia off at every pitch, 0.706 with lowest
  just above the 808 (measured by an earlier spike, not again with melodia
  below MIDI 52). legacy keeps only notes of 139 ms or more: on 32nd-note
  runs at 140 BPM its F1 was 0.36, melodic's 0.90 (4 clean renders through
  this server's worker). Without melodia a note needs an attack Basic Pitch
  sees: it found 3 of 16 pure sines with a 10 ms fade-in (chords 15), 13 of
  124 notes on synthetic 808 lines and 2 notes for the 336 of six 808 stems
  (MIDI 22-32; "No notes found" on five of them). So melodic decodes the
  model's output twice, melodia off from MIDI 52 (165 Hz) up and on below,
  and drops the low notes melodia reads an octave over a sounding one (a
  bass's second harmonic: 83 of the 106 extra notes on 3 saw bass lines).
  Onset F1, melodia off everywhere → melodic (3 renders each, 4 above MIDI
  52): 808 lines 0.18 → 0.86, with repeated notes 0.18 → 0.49, sine bass
  0.00 → 0.45, saw bass 0.84 → 0.82, sine bass at MIDI 40-55 0.34 → 0.82;
  plucks, runs, a lead and pads above MIDI 52 0.915 → 0.916, plucks and
  pads at MIDI 40-60 0.900 → 0.873. On the 808 stems F1 went from 0.006 to
  0.17: the notes have their pitch, but Basic Pitch starts them a median
  109 ms late (IQR 48–187 ms) at 29–52 Hz. When it finds no notes, the
  error suggests preset="chords".
• audio_to_midi(bends=true) lists the pitch bend of each bent note in
  cents, one value per model frame (11.6 ms). Basic Pitch 0.4.0 reads
  in-tune notes a third of a semitone sharp (+1 in 99.4–100 % of the frames
  of sines at 110, 220, 440 and 880 Hz); the values are corrected by that
  bin, after which in-tune and ±33-cent sines read 0 and ±33 cents. Live's
  API has no per-note pitch bend, so bends are returned, not written.
• audio_to_midi velocities follow each note's level
  (audio_analysis.note_levels): a band at the note's pitch, one semitone
  wide but at most 100 ms long, the loudest of its readings from 20 ms
  before the note's start to 60 ms after it, none past its end; the
  loudest note gets 127 and the others 127 × 10^(dB below it / 40). Rank
  correlation with the true level (Spearman, Basic Pitch on 4 clean
  synthetic renders each), against 127 × Basic Pitch's amplitude before:
  plucks 0.62 → 0.98, a legato lead 0.71 → 0.99, triad pads 0.27 → 0.95.
  On 3 driven 808 lines (MIDI 28-43, -24 to 0 dB, 82 notes) it is 0.98,
  and equal-level 808 lines get velocities 125–127. Centred on the note's
  start, a one-semitone window (0.51 s at MIDI 24) read the silence or the
  note before it: 0.95 and velocities 102–127, and equal-level 808 lines
  (MIDI 24-45) spread 3.6, 4.1 and 10.0 dB in quarters, 8ths and 16ths at
  140 BPM, against 0.08, 0.12 and 0.18 dB now. Below MIDI 53 the band is
  wider than a semitone, so a kick on an 808 note adds 0.9 dB on average.
  With reverb and detuned pad voices a similar measure ranked plucks at
  0.81–0.83, a lead at 0.57 and pads at 0.20. legacy keeps the old
  velocity.
• audio_to_midi places notes where the source clip plays them: a warped
  clip through Live's own warp (Clip.sample_to_beat_time, in batches), an
  unwarped one from its start marker; notes the clip never plays are left
  out. offset_beats moves every note, and timing in the result says
  what applied (or that Live could not map a warped clip). Checked against
  the fake Live, not yet in Live. lowest and highest also take Hz
  ("82.4 Hz").
• build_tracks loads a Simpler's or a drum pad's sample from a file
  (sample_path, absolute) with SimplerDevice.replace_sample (Live
  12.4+), builds those pads without the browser (insert_chain, in_note,
  a Simpler per chain) and reads each file back from its Simpler. By name,
  the browser's first hit could be another file with the same name. On
  older Live, or when Live refuses a file, the file's name goes to the
  browser (loaded_by_name says why), and the call fails if neither
  works. Simplers and pads take playback_mode (classic, one_shot,
  slicing) and start_marker (sample frames), pads choke_group (0–16),
  and signature sets the time signature first. Settings such as
  "Detune": "-12 ct" resolve by name as before. If Live refuses a call
  that builds pads by path (insert_chain, in_note,
  Chain.insert_device), the rack is deleted and every pad comes from the
  browser on a new one, as loaded_by_name says. sample may come with
  sample_path: it is the name the browser looks up when the path cannot
  be used, and what 2.2.0, which ignores sample_path and the new options
  without an error, loads. Everything goes through the Remote Script's
  generic LOM bridge: script 2.1.0 works unchanged. Measured in Live 12.4.6
  on 2026-10-19 with script 2.1.0, every option passed: files by path on a
  Simpler and on Drum Rack pads (each read back as the exact path),
  playback_mode classic and one_shot, start_marker (frames),
  choke_group, warping: false, "Detune": "-12 ct", signature "7/8",
  the browser fallback for a file Live refuses (on a Simpler and on one pad
  of a rack) and the refusals before anything is built; a 5-track build (an
  808 Simpler and drum pads, by path) took 7.7 s. Not verified in Live: the
  rack rebuilt by name (Live 12.4.6 refused none of those calls; the fake
  Live was made to refuse each), loading on Live before 12.4, and the
  slicing mode.

Changed

• audio_to_midi defaults to the melodic preset and to velocities from
  each note's level; preset="legacy" gives what it gave before.
• The tempo estimate of the audio analysis tools is found between frames: a
  parabola through the autocorrelation peak at one beat, then at 2 to 32
  beats, and it is reported to 0.01 BPM (0.1 before). Whole-frame lags are
  3.9 BPM apart at 140 BPM (44.1 kHz), and a 137 BPM click track read
  136.0; it reads 137.0 now. Click tracks from 70 to 175 BPM at 44.1 and
  48 kHz came within 0.01 BPM wherever the tempo octave was right; the
  octave choice is unchanged (150 BPM still reads 75).

Documentation

• capture_stems says its recordings are at the audio device's sample rate
  (often 48 kHz). Read as 44.1 kHz, a capture drifts by 9 %.
• CLAUDE.md: recordings' sample rate; Live keeps playing a sample it loaded
  after the file is overwritten on disk, until it restarts.
• CLAUDE.md: what Live 12.4.6 does with the Simplers and pads
  build_tracks makes, measured on 2026-10-19. replace_sample reads an
  overwritten file again, where a browser load keeps the old content.
  Transpose keeps whole semitones and Detune whole cents from −50 to +50,
  truncating numbers (display text such as "11.5 ct" rounds). A Simpler's
  level follows dB = 0.4 × A × (velocity − 70) − 0.096 over its Volume, A
  being "Vol < Vel" as a fraction (35 % on a Simpler made with
  insert_device, 45 % on a pad loaded through the browser). A choke
  group cuts with a 1 ms fade; "Snap", on by default, ends a sample
  10–30 ms early; Glide slides linearly in semitones over the Glide Time
  (one voice, overlapping notes); notes start within 0.04 ms of their
  time, and recordings add no offset.
• The compose-in-live skill says when preset="chords" fits and that Basic
  Pitch starts the lowest notes (under ~50 Hz) about 0.1 s late.
• CLAUDE.md: what Live 12.4.6 did in reference-dissector's live checks of
  2026-10-29, which built 8 plans and a song with 2.2.0's build_tracks
  and recorded every pad. One-Shot pads hit again start the new hit 1.000
  ms late (391 hits) and fade the old one out over 1 ms; the Classic 808
  Simpler releases at −1.683 dB/ms, restarts a note of its own pitch ~1 ms
  late, and its filter delays notes 0.13–0.46 ms; Simpler loses top end
  (more for 44.1 kHz files); a Volume sent as display text is held a step
  low; loading by file name took another file of the same name, and a
  name the browser never listed made the build stop half-made; cues land
  on the Arrangement's zoom-dependent grid.
• README and the ableton-setup skill: "Connection reset by peer" means the
  Remote Script already serves its 8 clients (one per MCP server).

Tests

• The fake Live, and so the simulator, reports Live 12.4.6 and has Drum
  Rack chains that a pad lists by their in_note (as DrumPad.chains
  does), SimplerDevice.replace_sample, Simpler's Detune and
  Clip.sample_to_beat_time. 28 new tests, 370 in all; 2 of them run the
  real Basic Pitch model, only with ABLETON_MCP_ML_TESTS=1.
• The companion tools' test gives them a connection that raises. It used the
  server's own, which tried localhost:9877, where a running Live listens,
  3 times per run (2.5 s; 0.45 s now). No test tries that port now (checked
  over the whole suite with Live open).

v2.2.0

Remote Script 2.1.0: reinstall it (ableton-mcp-install-script) and restart
Live. Checked in Live 12.4.3: the script loads without errors, recordings
start at start_beat, build_tracks, browser ids, clipping reports and
watch_live events work as before.

Removed

• The telemetry and dataset recorder inherited from upstream
  (MCP_Server/telemetry*.py, MCP_Server/dataset/, ~3,200 lines, the
  dataset extra with supabase). It was inactive in this fork (it needed an
  unshipped MCP_Server/config.py), but it read a Supabase service key from
  the environment and its six tools (set_dataset_consent, submit_intent,
  rate_last_action, prefer_candidate, reject_last_action,
  record_audition) were in the source. The server now reads no credentials.
• The Remote Script's passive listeners (on tempo, signature, playback and
  every track, clip slot and mixer) and drain_passive_events: only the
  dataset recorder read them, and they ran in Live all the time.

Fixed

• rack_control (info) crashed on a device that does not report
  visible_macro_count (an error dict was added to a number).

Tests

• Every tool is now called by at least one test (18 were not): the original
  full-profile tools run through the real Remote Script on the fake Live.

v2.1.2

Added

• build_tracks builds Drum Racks (a sample per pad, with settings), audio
  tracks with clips that start at the file's start (Live's warp analysis
  otherwise moves them to the first transient), mute, and Simplers with
  warping off. A sample named with its extension is found by exact name,
  waiting up to sample_wait_s for Live to list a file just written. Checked
  in Live 12.4.2: reference-dissector's PIRAMIT plan (an 808, six drum pads,
  675 notes, five stems) builds in 4.9 s, every setting as planned.

Fixed

• Recordings now really start at start_beat in Live 12.4.2. 2.1.1 moved the
  playhead before continue_playing, but continue_playing resumes where
  Live last stopped whatever the playhead says; recording now sets
  Song.start_time and calls start_playing (checked in Live: a section at
  beat 32 gives a 4.07 s file with its first hit at 0.9 ms).

v2.1.1

Remote Script 2.0.1: reinstall it (ableton-mcp-install-script) and restart
Live for the browser fix; everything else works with 2.0.0.

Added

• build_tracks builds MIDI tracks from a spec in one call: instrument (a
  device by name or a sample from the browser), settings, effects with
  Compressor sidechains (routed once every track exists), fader and pan, and
  Arrangement clips with their notes. Problems (a sidechain source Live does
  not offer, a setting that failed) are reported per track.
• capture_stems lists stems that reached full scale (clipped, and
  clipped_samples per stem); mix_doctor says so in its findings.

Fixed

• render_and_analyze, capture_stems and mix_doctor recorded from
  wherever the playhead was, not from start_beat: Live applies a playhead
  move only after the command that made it, and the move and
  continue_playing went in one batch. The playhead is now moved first and
  read back.
• audio_to_midi and the other ML jobs failed on results over 64 KB
  ("Separator is found, but chunk is longer than limit"), such as a dense
  3-minute stem's notes.
• audio_to_midi returned no notes when lowest was under A0 (27.5 Hz,
  Basic Pitch's lowest note); the floor is raised to it.
• Browser ids went stale during long searches ("Unknown id"): results shared
  the 4096-entry object registry and a walk of ~12,000 samples evicted them.
  They now have a registry of their own (Remote Script 2.0.1).
• edit_clip_notes says note times count from the clip's start (also for an
  Arrangement clip placed later); apply_device_chain says Gate's sidechain
  cannot be set from a script.

Changed

• manage_track lists its actions (create_midi, create_audio,
  create_return, delete, …) as an enum in its input schema, so a client
  or a registry sees what it can do without reading the description. An
  unknown action is refused with the list of valid ones; actions are now
  matched exactly (lowercase).

Fixed

• The .mcpb manifest summed up lom_call_method as "Call any LOM method,
  e.g.": a tool's summary no longer ends at "e.g." or "i.e.".

v2.1.0

The Remote Script is unchanged (2.0.0): no reinstall or Live restart needed.

Added

• Safety net. agent_changes lists what the assistant changed in this
  session and, with undo=N, takes back exactly the undo steps those changes
  made, counted the way Live 12.4 records them (parameter and other changes
  are separate steps; unchanged values, transport, selection, arm and solo
  make none). Parameter changes leave an automatic snapshot (the last 10),
  which still works after the user edited the set. The seven generators,
  transform_clip_notes and apply_midi_tool take preview=true. Deleting a
  track, scene, clip or device asks the user first in clients with MCP
  elicitation, in both protocol versions.

• Simulator. ableton-mcp-simulator runs the real Remote Script over an
  in-memory Live set (with an undo history that follows Live's rules), so the
  server and its tools work without Ableton; --state-file keeps the set as
  JSON.

• Eval suite (evals/): 32 claude plugin eval cases, each against a
  fresh simulated set, graded by deterministic checks of the resulting set
  (plus the reply where wording matters). A scripted correct solution of
  every case runs in the tests.

• Arrangement automation and transitions. write_arrangement_automation
  puts automation over a stretch of the Arrangement on MIDI tracks at once
  (the notes there go into a Session clip with the envelope, laid back over
  the stretch: Live 12.4 draws envelopes only on Session clips, and a copy
  keeps them). record_automation records track automation on any track,
  the main one included, by moving the parameter while the Arrangement
  records. design_transition builds risers, reversed crashes, drum fills,
  gaps before a drop and high-pass sweeps. All verified on Live 12.4,
  including undoing each with agent_changes.

• Audio generation on this computer. generate_audio makes a sound or
  a loop from a text prompt with MusicGen small (in its own environment,
  like the other models) and puts it in the set: a clip on an audio track,
  a Drum Rack pad, or a Simpler. The set's tempo and key go into the prompt;
  the files land in the User Library. MusicGen's weights are CC-BY-NC 4.0,
  which the tool says (non-commercial use).

• Project memory. memory keeps notes about a song inside the Live set
  (Song.set_data, saved with the set, no undo step) and the user's
  preferences in ~/.ableton-mcp/memory.json; a keyed note replaces the
  earlier one. get_session_overview shows them, so a later session (or
  another client) starts from what was decided.

• MIDI Tools in Live's clip view. install_midi_tools adds three Max for
  Live MIDI Tools that run inside Live, without the server: Usul and Makam
  Melody (Generate) and Makam Snap (Transform). Their JavaScript is a port
  of the server's generators (the usuls match note for note) and lives in
  the device. Verified in Live 12.4 with Max 9.1.4.

Changed

• Less context per session and per call. The core profile's tool
  descriptions and schemas take about 15k tokens instead of 22k: schemas lose
  pydantic's titles and null defaults, long descriptions are tighter, and six
  original tools that newer ones replace (load_instrument_or_effect,
  get_browser_tree, get_browser_items_at_path, load_drum_kit,
  get_remote_script_info, get_arrangement_clips) moved to the full
  profile. Results are compact JSON (a quarter to a half smaller), and notes
  come as rows under note_columns (an 8-bar melody: 885 characters instead
  of about 4,100).
• The server takes Live's port from the session file the Remote Script
  writes (ABLETON_PORT still wins).
• Makam melodies arrive on the karar through the makam's first tetrachord
  (G F# Eb D in Hicaz on D), so its characteristic intervals are always
  heard; before, one 4-bar Hicaz melody in five never played its F#.

Fixed

• While a client watched a resource (Claude Code does), the background
  watcher's 2-second long polls held the connection the tools use, so a tool
  call could wait up to 2 s per command. The watcher now has a connection of
  its own. Found by the eval suite (a recorded fade lost whole seconds).
• A server started before Live refused create_locator and the other tools
  that check the Remote Script's capabilities until it was restarted: the
  capabilities are now taken from every connection's hello.

v2.0.0

Breaking

• Remote Script 2.0. Commands run on Live's main thread (about 10 ms
  each instead of ~410 ms), one JSON object per line with request ids, and a
  per-launch session token. Reinstall the script (ask the assistant to run
  install_remote_script, or ableton-mcp-install-script) and restart Live;
  servers older than 2.0 cannot talk to it.
• MCP Python SDK 2.x (mcp>=2.3), Python 3.10 or newer.
• Tool profiles. The default core profile has one tool per job; the
  original tools that newer ones replace need ABLETON_MCP_PROFILE=full.

Added

• Working with Live: one undo step per tool call; doctor and
  install_remote_script; Live's changes pushed to subscribed clients
  (ableton://events, ableton://session/overview, both MCP protocol
  versions); progress notifications for long tools.
• Writing music: generators that follow the set's meter and Key & Scale
  (chords, drums, bass, arpeggios, melodies, Euclidean rhythms); harmony
  analysis; Live 12 MIDI Tools equivalents; Turkish makams with .ascl
  tunings and usuls; build_arrangement; MIDI file import and export; drum
  kits from one-shots; browser search, loading and preview.
• Mixing and listening: measure_levels, render_and_analyze,
  capture_stems, mix_doctor (masking between parts), compare_to_reference
  with EQ Eight settings, snapshots with A/B, ready-made device chains.
• Panels (MCP Apps) for 22 tools: mixer and Session grid, meters, masking
  heat map, reference curve, snapshots, piano roll, drum pads, makam keyboard.
• Live Set files: read_live_set (and comparing with a backup),
  find_live_sets, import_from_live_set, without opening Live.
• Machine learning, each in its own environment set up by uv on first
  use: separate_stems (Demucs), audio_to_midi (Basic Pitch),
  find_similar_sounds (CLAP).
• Max for Live analyzer: install_analyzer puts a generated device on a
  track; read_analyzer reports octave bands, loudness, correlation and width
  while the set plays.
• Extensions SDK add-on (extension/, Live 12.4.5+ Suite beta; not yet
  run inside Live): get_live_selection, confirm_in_live,
  render_track_audio.
• Distribution: a Claude Desktop extension (.mcpb), a Claude Code plugin
  with five skills, CI, and GitHub releases from version tags.