convez.ai
MCP ServerDesenvolvimento WebAvançado
ehm-93

mixdown – Síntese de Efeitos Sonoros por IA

ehm-93/audio-mcp

Design e síntese de SFX para jogos via IA, sem ouvir

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que permite a um modelo de linguagem projetar, analisar e exportar efeitos sonoros para jogos como receitas de síntese determinística armazenadas em JSON. O servidor renderiza áudio, retorna medições e espectrogramas que o modelo pode ler, valida defeitos comuns, gera variantes e emite uma página de audição para validação humana final.

Instalação

$npx convez-ai ehm-93/audio-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar efeitos sonoros procedurais para games diretamente do Claude
  • Analisar e otimizar receitas de síntese via espectrogramas e métricas
  • Iterar famílias de variantes de som com feedback visual e medições

Veredito

Ferramenta especializada e sofisticada para audio-design procedimental em gamedev. Exige familiaridade com síntese sonora e fluxo de produção, mas resolve um nicho real onde IA e DSP se encontram de forma prática e determinística.

README

A tradução para o português deste README ainda está na fila.

mixdown

A local MCP server that lets a language model design, edit, analyze, and batch-export game sound effects without being able to hear them. Sounds are deterministic synthesis recipes stored as JSON; the server renders them to audio, returns measurements and spectrogram images the model can read, lints for common defects, generates variant families, and emits an audition page so a human makes the final call by ear.

See SPEC.md for the full design.

Setup

Requires Node 20+. No native compile step — wav/ogg/flac IO is pure JS + wasm, spectrograms use prebuilt binaries.

npm install
npm run build

Add to your MCP client config (e.g. Claude Desktop or .mcp.json):

{
  "mcpServers": {
    "mixdown": {
      "command": "node",
      "args": ["/path/to/audio-mcp/dist/index.js", "/path/to/your/sfx-workspace"]
    }
  }
}

The workspace argument defaults to $MIXDOWN_WORKSPACE, then the current directory. The server creates the layout on first run:

workspace/
  project.json          shared settings and palette
  sounds/               one JSON recipe per sound
  refs/                 imported reference audio (wav, ogg, flac)
  renders/              cache, content-addressed, safe to delete
  exports/              game-ready files
  audition/index.html   generated listening page
  history/              recipe version snapshots

Tools

toolpurpose
list_soundsevery recipe with version, duration, variant count, worst lint severity
read_recipe / write_recipe / patch_recipeCRUD with versioning; patch is RFC 7386 merge patch
patch_layermerge patch one layer by id without resending the layers array
list_presetsevery enum the server accepts — call once per session
renderrender (cached by content hash) → analysis bundle, lint findings, spectrogram/waveform PNGs
previewstateless render of an inline recipe — same analysis as render, zero workspace writes; for parallel exploration (subagents iterate, commit the winner with write_recipe)
comparemetric deltas, per-band energy deltas, stacked spectrograms on identical axes; each side is a soundref or an inline recipe
lintfindings only, served from the render cache
make_variantswrite the variants block, render the family, return a table + strip image
import_referencecopy outside audio into refs/ and analyze it
freezebounce a rendered sound into refs/ for reuse as a sample/granular/IR source
audition_pageregenerate the human listening page
exportloudness-normalized ogg/wav with naming template; refuses on lint errors unless forced
history / checkpoint / revert / diffversion management

v0.2 recipe features: per-layer LFO (gain/pitch/cutoff with sweepable rate), breakpoint pitch envelopes, per-segment envelope curves, master fades, automatic delay/reverb tail ring-out, per-recipe lint allow-list, hidden flag. See the v0.2 section in SPEC.md.

Development

npm test                          # vitest: DSP, store, end-to-end
node --import tsx scripts/smoke.ts   # protocol-level smoke test over stdio (build first)
node --import tsx scripts/preview.ts # render sample PNGs to tmp-preview/

Determinism: a recipe plus an engine version renders byte-identical output on the same Node major; the engine version embeds the Node major (e.g. 0.1.0-node26) and is part of the render cache key.

Ver o original no GitHub

Bom para

Relacionados