Skip to content

Latest commit

 

History

History
261 lines (199 loc) · 19.8 KB

File metadata and controls

261 lines (199 loc) · 19.8 KB

LivePilot Manual

An agentic production system for Ableton Live 12. 472 tools. 57 domains. Device atlas. Sample intelligence. Auto-composition. Spectral perception. Technique memory. Creative intelligence.


What LivePilot Is

LivePilot is not a tool collection with an AI wrapper. It is a production system — three perception layers feed into 472 tools, which are orchestrated by a dozen creative engines that understand song identity, learn your taste, diagnose session problems, and generate real musical options.

The difference: a tool collection executes "set volume to -6dB." LivePilot understands that turning down the drums might kill the groove that defines the track, suggests three genuinely different ways to create space instead, lets you preview each one, and remembers which approach you preferred.


Architecture

AI Client  ──MCP──►  FastMCP Server  ──TCP/9878──►  Remote Script (inside Ableton)
                        (validates)                    (executes on main thread)
                            │
                            ├── Device Atlas (5264 devices, 120 enriched, 7 indexes)
                            ├── User Corpus (~/.livepilot/atlas-overlays/)   [v1.23.4+]
                            │      ├── user/         your detected plugins
                            │      ├── m4l-devices/  your .amxd library
                            │      ├── packs/        factory + your packs
                            │      └── elektron/     hardware overlays
                            ├── M4L Analyzer ──UDP/OSC──► LivePilot_Analyzer.amxd
                            └── Technique Memory (~/.livepilot/memory/)

The atlas resolves device names and browser URIs — the AI never hallucinates a preset. 641 devices are indexed by pack (Core Library + explicit-pack assignments) so "what's in Drone Lab?" is an instant lookup. Reverse-index device_techniques_index.json cross-references 146 techniques across 58 devices (the atlas_techniques_for_device tool). The user corpus (v1.23.4+) extends the atlas with whatever's installed on YOUR machine — third-party plugins, custom Max devices, your .adg racks. Loaded by every reasoning tool with a 50/50 result-budget split alongside the factory atlas, so "sound like Burial" can route to your CHOWTapeModel + Splice PHOTEK pack instead of just Ableton's stock saturator. See User Corpus Guide. The analyzer feeds back spectral data from the master bus so the AI hears its own changes — 9 frequency bands (sub_low / sub / low / low_mid / mid / high_mid / high / presence / air); the sub_low band (20-60 Hz) separates kick fundamental from DC rumble. From v1.20.3 the analyzer is auto-loaded via ensure_analyzer_on_master — the Creative Director skill calls this at the top of every turn's Phase 1 ground read. The memory persists production decisions across sessions as searchable, replayable data structures.

All 472 tools execute as deterministic LOM calls on Ableton's main thread. Live-session mutations (clips, devices, mixer, arrangement) route through Ableton's undo stack; side effects that touch state outside the Live project — Splice downloads, memory/ledger writes, installer actions, atlas scans, filesystem writes — persist beyond undo.


The Four Layers

Layer 1: Deterministic Tools

The foundation. Direct control over every aspect of Ableton Live through the Live Object Model: transport, tracks, clips, notes, devices, scenes, mixing, arrangement, browser, automation, scales, grooves, take-lanes, follow actions, MIDI tool generators. Also includes music theory (Krumhansl-Schmuckler key detection, neo-Riemannian harmony, species counterpoint), generative algorithms (Euclidean rhythm, tintinnabuli, phase shift, additive process), and MIDI I/O.

These tools do exactly what you ask. No interpretation, no judgment. They are the building blocks.

Layer 2: Perception

The M4L Analyzer on the master track gives LivePilot ears: 9-band FFT spectrum (v1.17+), true RMS/peak metering, Krumhansl-Schmuckler key detection, pitch tracking, FluCoMa mel-spectrum + chroma + onset detection + spectral shape. Plus deep LOM access for hidden parameters, automation state, Simpler internals, and warp markers. Offline perception adds loudness analysis (integrated LUFS, LRA), spectral analysis, and reference comparison. A second, capture-based path (v1.28+) runs the same evaluation-stack feature extraction offline against rendered WAV/AIFF files — listen_capture for a full report, listen_ab for a before/after delta — producing canonical snapshots evaluate_move can score directly, independent of the M4L bridge.

Perception closes the feedback loop. Without it, the AI is blind to the result of its own changes. With it, the AI can verify that a mix move actually reduced masking, that a filter sweep landed at the right frequency, that the overall loudness is broadcast-ready.

Layer 3: Creative Intelligence

Sits on top of the tools and perception, adding musical judgment. This is what makes LivePilot agentic:

Understanding the song:

  • SongBrain builds a real-time model of identity, sacred elements, section purposes, and energy arc
  • Composition Engine detects motifs, infers emotional arcs, plans structural moves
  • Hook Hunter finds the most salient musical idea and tracks whether it's being developed or neglected

Understanding the producer:

  • Taste Graph / Agent OS learns move family preferences, device affinities, novelty tolerance, and dimension avoidances across sessions
  • Session Continuity tracks creative threads, turn resolutions, and the session story
  • Technique Memory stores and recalls production decisions by mood, genre, and texture

Making musical decisions:

  • Semantic Moves express high-level intent ("add contrast," "tighten the low end") as executable tool sequences with risk levels and protection thresholds
  • Wonder Mode diagnoses stuck sessions, generates genuinely distinct rescue options, and lets you preview before committing
  • Preview Studio renders variants using Ableton's undo system — hear each option, compare, then choose
  • Creative Constraints activate style gates and render reference-inspired variants

Evaluating results:

  • Mix Engine runs critic-driven analysis — masking, headroom, stereo, dynamics — and plans corrective moves
  • Sound Design Engine analyzes patches for static timbre, missing modulation, and weak transients
  • Synthesis Brain extracts timbre fingerprints + proposes synth branches
  • Transition / Reference / Translation Engines score transitions, distill principles from reference tracks, translate across domains
  • Evaluation Loop enforces measure-before, act, measure-after discipline on every creative move

Layer 4: Personal Library Awareness [v1.23.4+]

The first three layers operate on Ableton's shipped library. Layer 4 makes the brain aware of your specific machine — every third-party plugin, every Max for Live device, every custom rack you've saved.

Why this layer exists. Most music AI tools have a fixed knowledge cutoff. LivePilot's first three layers give it generalized musical judgment, but until v1.23.4 they could only reach for Ableton-shipped gear. If you spent $500 on Valhalla plugins, $200 on a Cem Olcay MIDI suite, and built a personal library of .adg racks, none of that was visible to the brain. Same query → same Ableton-stock recommendation, regardless of what's actually on your hard drive.

What it does. The User Corpus runs a 4-phase pipeline (DETECT → CANONICALIZE → RESEARCH → SYNTHESIZE) that turns the contents of your filesystem into AI-queryable knowledge:

Phase Tool What it produces
1. Detect corpus_detect_plugins Inventory of installed VST3/AU/AUv3/VST2/AAX/CLAP/LV2 plugins. Path-walks /Library/Audio/Plug-Ins/{VST3,Components} + runs auval -a for AUv3/Mac Catalyst coverage
2. Canonicalize corpus_canonicalize_plugins Dedupes by vendor+name, prefers VST3 over AU, strips vendor suffix variants ("Valhalla DSP, LLC" = "Valhalladsp" = "Valhalla DSP")
2.5. Cluster corpus_cluster_plugins Groups by vendor for batched research dispatch — vendors with ≥2 plugins share one WebSearch
3. Research corpus_research_targets Emits a structured WebSearch task packet the agent fulfills
4. Synthesize corpus_emit_synthesis_briefs Sonnet-subagent dispatch — writes per-plugin identity.yaml with sonic_fingerprint / reach_for / avoid / key_techniques / parameter_glossary / genre_affinity / producer_anchors

How it integrates with the brain. The output overlays land at ~/.livepilot/atlas-overlays/<namespace>/ across four namespaces (user/ for plugins, m4l-devices/ for Max devices, packs/ for factory + custom packs, elektron/ for hardware). Every reasoning tool — atlas_search, atlas_chain_suggest, atlas_macro_fingerprint, atlas_describe_chain — consults the overlay alongside the factory atlas with a 50/50 result-budget split. Each result is tagged source: factory_atlas | user_overlay:<namespace>, so the agent always knows whether a recommendation is Ableton stock or yours.

The versatility argument in one example. A new user asks "make me a J Dilla beat":

Without User Corpus With User Corpus
Operator FM marimba (default preset) Koala Sampler (your SP-404 emulator)
Live's stock Drum Rack samples Splice PHOTEK Vol.1 pack you bought
Saturator (Tape mode) for warmth CHOWTapeModel (your peer-reviewed tape model)
Operator-based bass (default) Moog Model D (your subby analog bass)

Same query, different brain. The closer the brain matches your library, the more useful every recommendation becomes.

Quick start. From any MCP client:

corpus_setup_wizard           # one-shot — runs the full pipeline
# OR fine-grained:
corpus_detect_plugins use_auval=true
corpus_canonicalize_plugins
corpus_cluster_plugins
corpus_research_targets
corpus_emit_synthesis_briefs

See the full User Corpus Guide for end-to-end walkthrough, Plugin Knowledge Engine doc for engine internals, and Extension API for the overlay file format.


Domain Map

The table below is auto-generated from the live tool registry — do not hand-edit the region between the DOMAIN_MAP markers. Regenerate with python3 scripts/sync_metadata.py --fix (or directly via python3 scripts/generate_tool_catalog.py --domain-map); --check fails CI if this block drifts from the registry.

All 472 tools across 57 domains, in source-truth per-domain counts:

Core Ableton Control (Layer 1 — 245 tools)

Domain # Scope
Devices 42 Load by name/URI, insert native (12.3+), params, racks, chains, drum pads, plugins, presets, wavetable mod matrix, replace_sample (12.4+), add_drum_rack_pad, verify_device_alive (v1.16+)
Arrangement 21 Timeline editing, arrangement notes, native clips (12.1.10+), cue points, recording, capture, set_arrangement_automation_via_session_record (v1.17+)
Tracks 21 Create MIDI/audio/return, delete, duplicate, arm, mute, solo, routing, sends, monitoring
Transport 21 Playback, tempo, time sig, loop, metronome, undo/redo, diagnostics, capture MIDI
Memory 18 Save, recall, replay, session memory, list/favorite/delete, update
Clips 16 Create, delete, duplicate, fire, stop, loop, launch mode, warp mode, pitch, color
Scenes 12 Create, delete, duplicate, fire, name, color, per-scene tempo, follow actions
Mixing 11 Volume, pan, sends, routing, meters, return tracks, mix snapshot
Automation 9 Clip envelopes, 16 curve types, 15 recipes, spectral suggestions
Composition 9 Section analysis, motif detection, emotional arc, form planning
Follow Actions 8 Clip + scene follow actions, presets, type listing
Notes 8 Add/get/remove/modify MIDI, transpose, duplicate, per-note probability
Scales 8 Clip scales, song scales, scale modes, list available scales
Grooves 7 Groove templates, per-clip groove, groove amount, groove params
Theory 7 Harmony analysis, Roman numerals, scales, countermelody, transposition
Take Lanes 6 Create, name, list, per-lane clips
Generative 5 Euclidean rhythm, tintinnabuli, phase shift, additive process
Browser 4 Search library, browse tree, load items
Harmony 4 Tonnetz navigation, voice leading, neo-Riemannian classification
MIDI I/O 4 Export/import .mid, offline analysis, piano roll extraction
MidiTool 4 Device install, generator registration, per-clip target mapping

Perception (Layer 2 — 51 tools)

Domain # Scope
Analyzer 38 9-band spectrum (v1.17+), RMS, key detection, Simpler ops, warp markers, capture, FluCoMa mel/chroma/onset [M4L]
Listening 5 Offline perceptual analysis of master-bus captures — listen_capture full report, listen_ab before/after comparison with evaluate_move-ready snapshots (v1.28+)
Perception 4 Offline loudness, spectral analysis, reference comparison, metadata
Diagnostics 3 Device/session health verification, test-note fire-and-forget
Evaluation 1 Before/after evaluation with structured scoring

Creative Intelligence (Layer 3 — 162 tools)

Domain # Scope
Sample Engine 23 Multi-source search (Splice gRPC + browser + filesystem), Splice catalog hunt, downloads, previews, pack info, collections, presets, describe-a-sound (LIVE), variations (LIVE), http-diagnose (v1.17+)
Atlas 22 Search 5264 devices, suggest by intent, chain building, comparison, library scan, atlas_pack_info, atlas_describe_chain (free-text), atlas_techniques_for_device (reverse-lookup), atlas_macro_fingerprint (preset similarity), Pack-Atlas cross-pack tools (atlas_transplant, atlas_extract_chain, atlas_cross_pack_chain, atlas_pack_aware_compose, atlas_demo_story), hybrid-surface Layer B (atlas_explore, atlas_audition, atlas_substitute) — all v1.17+ unless noted; plus extension_atlas_search / extension_atlas_get / extension_atlas_list for user-local overlays (v1.23.0+)
Agent OS 9 Session kernel, action ledger, capability state, routing, goal vectors, taste, turn budget
Composer 9 Prompt → multi-layer composition plan, sample augmentation, plan preview, branches, develop-mode apply, hybrid Ableton-Knowledge consultation
Hook Hunter 9 Hook detection, salience scoring, neglect detection, phrase impact
Runtime 7 Session kernel building, world model, capability, safety, resume intent, Live 12.4 probe-first tools (probe_link_audio, probe_stem_workflow)
Session Continuity 7 Creative threads, turn resolution, session story, anti-preferences
Mix Engine 6 Critic-driven mix analysis, issue detection, move planning
Musical Intelligence 6 Phrase arc, impact scoring, comparison, rendering, grid analysis, snapshot
Creative Constraints 5 Constraint activation, reference-inspired variants
Experiment 5 Create, run, compare, commit, discard A/B experiment branches
Preview Studio 5 Variant creation, preview rendering, comparison, commit, discard
Semantic Moves 4 Move listing, preview, application, next-best-move proposal
Sound Design 4 Patch analysis, modulation planning, timbre scoring
Device Forge 3 Generate M4L devices from gen~ templates, install to browser
Grader 3 Rubric-based session evaluation — list rubrics, run one, run all
Performance Engine 3 Safety-constrained suggestions, safe moves, scene handoff
Reference Engine 3 Reference profiling, principle distillation, gap analysis
Research 3 Technique research, style tactics
Song Brain 3 Identity inference, sacred elements, drift monitoring
Stuckness Detector 3 Momentum analysis, rescue classification, rescue workflows
Synthesis Brain 3 Synth patch analysis, branch proposals, timbre fingerprint extraction
Transition Engine 3 Transition classification, scoring, archetype planning
Wonder Mode 3 Diagnosis-driven variants, taste-aware ranking, session discard
Creative Director 2 Brief-compliance checking for a proposed tool call, hybrid-brief compilation from multiple concept packets
Motif 2 Motif graph, motif transformation
Planner 2 Gesture planning, arrangement planning
Project Brain 2 Project-level analysis, section purpose inference
Translation Engine 2 Cross-domain translation, issue detection
Audit 1 Single-call §5 layer-precision audit (timbre + sequence + stereo + masking) for one track

Personal Library Awareness (Layer 4 — 14 tools)

Domain # Scope
User Corpus 14 Plugin Knowledge Engine 4-phase pipeline. corpus_setup_wizard, corpus_init, corpus_status, corpus_list_scanners, corpus_add_source, corpus_remove_source, corpus_scan (file-level), corpus_detect_plugins (auval-aware), corpus_canonicalize_plugins (VST3-preferred dedup), corpus_cluster_plugins (vendor-batched research), corpus_trim_plugin_identity (deprioritization), corpus_discover_manuals, corpus_research_targets, corpus_emit_synthesis_briefs

Chapters

Understanding the System

Chapter What's inside
The Intelligence Layer How the engines connect — conductor, kernel, moves, preview, evaluation
Device Atlas 5264 devices indexed — search, suggest, chain building, comparison, pack browsing
User Corpus Guide [v1.23.4+] Build a personal atlas from your plugins, racks, Max devices, samples — same MCP tools as factory atlas
Plugin Knowledge Engine [v1.23.4+] 4-phase detect → canonicalize → research → synthesize pipeline, AUv3 detection via auval, VST3-preferred dedup
Extension API Drop YAML overlays at ~/.livepilot/atlas-overlays/<namespace>/ — survives npm updates
Samples & Slicing 3-source search, Splice describe-a-sound, variations, fitness critics
Automation 16 curve types, 15 recipes, spectral suggestions, two-phase arrangement-record workaround
Composition & Arrangement Composer, section analysis, arrangement planning with sample roles

Production Guides

Chapter What's inside
Workflows Step-by-step: beats, session setup, sound design, arrangement, mixing
MIDI Guide Drum patterns, scales, chords, humanization techniques
Sound Design Instruments, effects, parameter recipes, device chains
Mixing Gain staging, EQ, compression, sends, stereo width

Reference

Chapter What's inside
Tool Catalog Every tool organized by domain (auto-generated from source)
Tool Reference In-depth parameter docs, ranges, and usage notes for the most-used tools
Troubleshooting Connection issues, common errors, diagnostics
Splice Endpoint Capture mitmproxy runbook for capturing additional Splice GraphQL operations

Contributing

Chapter What's inside
Dev Install Run LivePilot from a local checkout — venv setup, local MCP config, iterate-without-publish workflow, test suite, sync_metadata drift check

Next: The Intelligence Layer