N0zoM1z0/vocaloid-mcp

npm install && npm run build && npm test (Node.js 20+; needs npm TypeScript toolchain, xmllint, FFmpeg, FluidSynth + a compatible SoundFont).
Start the server with VOCALOID_MCP_PROFILE=production npm start.
Point an MCP client (stdio) at node /absolute/path/to/vocaloid/dist/src/index.js with env VOCALOID_MCP_PROFILE=production.
Actually singing requires your own legitimate Windows VOCALOID3 Editor + licensed voicebank; the repo contains no VOCALOID binaries, voicebanks, or final WAVs.

Usage

• •Begin from an empty timeline or a creative brief (copy intents/template.intent.json to your song, specify tempo BPM min/max/preferred, and run npm run --silent compose:prompt -- --intent=...).
• •Work the pipeline: intent -> theory/form/harmony/melody -> Japanese note allocation -> native IA delivery -> arrangement -> VSQX + hash-bound creative record.
• •Render the selected locally licensed singer through the original VOCALOID3 Editor; produce stems, deterministic effect chains, mixes, A/B pairs, and masters.
• •Measure LUFS/LRA/true peak/clipping/balance and score-vs-render pitch/timing/release/vibrato evidence.
• •Transfer to fresh Codex sessions with one fixed prompt plus one strict per-song intent file.

Key features

• ✓Creates native VSQ3/VSQ4 from zero; validates with Yamaha's schemas and a hash-pinned native Vsq3.dll loader probe.
• ✓Japanese lyric authoring with explicit kana readings, phonemes, mora allocation, rests, melisma, small-tsu timing.
• ✓Deterministic theory/arrangement diagnostics guardrail (structural errors vs musical risks vs style observations) that never composes the next note.
• ✓Canonical *.song.json manifest with stable IDs, dry-runs, bounded edits, history, separate content/composition/mix hashes, and an append-only creative journal.
• ✓Evidence ladder (declared/materialized/rendered/heard) — no technical QC ever self-signs a human keep decision; 102 tests (100 pass, 2 env-skipped); MIT.

Best for

Agents doing real VOCALOID3/4 composing/production work where genuine VSQX artifacts + honest failure evidence matter more than a one-button song generator.

Caveats

• ⚠Not a MIDI-to-VSQX converter and NOT a one-button hit-song generator; the licensed original Editor is the synthesis authority.
• ⚠Singing requires your own legitimate Windows VOCALOID3 Editor + installed/licensed voicebank; local single-user desktop bridge, not a remote multi-tenant service.
• ⚠Explicit non-claim: passing engineering gates does not mean a song is musically successful (their 190 BPM IA_ROCKS test still 'did not feel like rock').
• ⚠User-supplied assets (voicebanks, lyrics, corpus material) remain under their owners' licenses; project is a completed research prototype, not production-ready for arbitrary hosts.