mixdown – Síntese de Efeitos Sonoros por IA
ehm-93/audio-mcp
Design e síntese de SFX para jogos via IA, sem ouvir
Ver no GitHubMCP 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-mcp1 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
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
| tool | purpose |
|---|---|
list_sounds | every recipe with version, duration, variant count, worst lint severity |
read_recipe / write_recipe / patch_recipe | CRUD with versioning; patch is RFC 7386 merge patch |
patch_layer | merge patch one layer by id without resending the layers array |
list_presets | every enum the server accepts — call once per session |
render | render (cached by content hash) → analysis bundle, lint findings, spectrogram/waveform PNGs |
preview | stateless render of an inline recipe — same analysis as render, zero workspace writes; for parallel exploration (subagents iterate, commit the winner with write_recipe) |
compare | metric deltas, per-band energy deltas, stacked spectrograms on identical axes; each side is a soundref or an inline recipe |
lint | findings only, served from the render cache |
make_variants | write the variants block, render the family, return a table + strip image |
import_reference | copy outside audio into refs/ and analyze it |
freeze | bounce a rendered sound into refs/ for reuse as a sample/granular/IR source |
audition_page | regenerate the human listening page |
export | loudness-normalized ogg/wav with naming template; refuses on lint errors unless forced |
history / checkpoint / revert / diff | version 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.