Check whether REAPER is running and the Orpheus bridge loop is alive.

Call this first. Round-trips a real ping through the bridge and returns REAPER's version + latency, or a clear "not listening" message (rather than a mystery hang) if the in-REAPER script isn't running.

Tempo, time signature, length, track count, and transport state.

The typed track tree: name, vol/pan, mute/solo, role, FX-chain names, item count.

Read a take's MIDI notes (pitch, start/duration in BEATS, velocity) + CC.

Decode an FX chain's parameters BY NAME (e.g. compressor ratio), not as opaque 0–1.

Set the project tempo in BPM (downstream: seconds_per_beat = 60 / bpm).

Set the project time signature.

Transport control: 'play' | 'stop' | 'record', via Main_OnCommand action IDs.

Insert a named track and return its stable GUID (one composite bridge call).

Set fader/pan. Accepts fuzzy values ('-6dB', '+3', '50%', 'L50', 'center').

Add an FX by fuzzy name, VALIDATED against the installed-plugin inventory.

On a no-match returns a graceful "not found — here's how to install it" message rather than silently loading the wrong plugin (the bug all three live servers have). Orpheus never auto-installs plugins.

Set an FX parameter with name→index resolution (so 'ratio' works, not just index 7).

Batch-write notes into a take, PPQ/tempo-correct, in ONE bridge round-trip.

Notes are in beats (pitch, start_beat, duration_beats, velocity). Batched because per-note round-trips would hit the ~10 ops/sec file-IPC ceiling. Capped per call (see BridgeClient.MAX_NOTES_PER_CALL) so REAPER's audio thread never stalls.

Create a MIDI item/take over a bar range; returns a stable take identity.

Quantize a take's notes to a grid (cleanup before/after analysis).

Transpose a take's notes — the APPLY verb for key/mode retargeting.

MIDI notes for a key/mode so generated notes stay diatonic.

A genre-typical diatonic progression as Roman numerals + concrete MIDI.

Snap a proposed note set to the nearest in-key pitches before writing.

A style's typical progressions / scales / BPM range / instruments / drum rhythms. The RECOMMEND-side lookup.

Detect key/scale/chords/Roman numerals/cadences.

Exports the project's MIDI, filters drum tracks (Krumhansl detection is garbage on beats), and runs music21. Always returns a confidence + alternatives; on drum-only or non-chordal material it hedges rather than inventing harmony.

Compute swing%% and tightness from raw MIDI PPQ onset deviation off the grid. No existing DAW-MCP server provides this — a genuine Orpheus differentiator.

Render stems/master to WAV and compute the post-FX sonic fingerprint: 3-band energy, spectral centroid, LUFS, true peak, crest factor, stereo width. (Sharper with the optional [analysis] extra / librosa.)

Fuse symbolic + audio analysis into one CompositionSpec(current) + a plain-English report. The contract handed to recommend_changes.

List cached target fingerprints (e.g. 'classical', 'hiphop', 'dominic-fike-sunburn'). Each is the analysis pipeline run over 3–5 per-era reference tracks.

"Why does this sound like X?" — run feature thresholds over the current project and return human-readable reasons (each matched threshold = one reason).

This is the v0.1 headline capability: understanding without transformation.

Diff the current project against a target fingerprint and return an EditPlan — a list of ProposedEdit{target, action, reason, params} across the v1 dimensions (tempo, key, harmony, instrumentation, mastering-match), each annotated with a human-readable reason. Read-only: surface these reasons to the user for approval, THEN call apply_changes. Includes honest caveats (detection confidence, taste calls).

Execute an approved EditPlan: symbolic edits (transpose, rewrite progression MIDI, swap/add FX, retarget tempo, write grooved patterns), all wrapped in one Undo_BeginBlock/EndBlock so a single Ctrl+Z reverts everything.

Pass back the exact plan the user approved — do not re-derive it here.

The spectral half: render the master → Matchering against the fingerprint's reference WAV → bake the matched curve onto a master-bus ReaEQ + limiter.

Headless render to WAV/MP3 (RENDER_FILE + format/bounds + Main_OnCommand(41824)).

Render per-track stems by sequential solo — feeds analyze_audio_character.

Render, then measure LUFS/spectrum in one call so the agent self-checks ('loud enough? too bright?') and iterates without a human ear.