sound-forge
Local text-to-audio game sound generator for Claude,Codex and any MCP client: sound effects, seamless amience loops, stingers, looping music and consistent haracter voices as game-ready .wav/.ogg on your own PU, via ComfyUI (Stable Audio 3, ACE-Step 1.5, Qwen3TTS). MCP server + agent skill.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/noamcelermajer--comfyui-text-to-audio-mcp)HTML
<a href="/listings/noamcelermajer--comfyui-text-to-audio-mcp"><img src="/api/listings/noamcelermajer--comfyui-text-to-audio-mcp/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
The easiest way: paste [SETUP_PROMPT.md](SETUP_PROMPT.md) into Claude Code or Codex. By hand:
git clone https://github.com/Noamcelermajer/comfyui-text-to-audio-mcp SoundForge cd SoundForge python forge.py setup # node pack + models for this GPU (--dry-run to preview, --all for the optional ones) python forge.py doctor # every profile should be "ready": true python forge.py connect # registers the MCP server and the skill for Claude and Codex
Sound Forge finds ComfyUI next to it or through `SOUNDFORGE_COMFY`, starts it when needed, and hands over to ComfyUI's Python for the audio libraries. Full unattended instructions for agents: [AGENTS.md](AGENTS.md).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 11 lines
- {
- "name": "sound-forge",
- "owner": {"name": "Noamcelermajer"},
- "plugins": [
- {
- "name": "sound-forge",
- "source": "./",
- "description": "Local game audio generation (SFX, ambience, music, voice) on ComfyUI, as an MCP server plus a skill."
- }
- ]
- }
README.md · README · 146 lines
- # Sound Forge - local text-to-audio game sounds for Claude, Codex and any MCP client
- **Turn a sentence into a game-ready sound on your own GPU.** Free, private and offline:
- sound effects, seamless ambience loops, music stingers, looping background music and
- character voice lines, delivered as trimmed, loudness-normalised `.wav` / `.ogg` files and
- driven by your coding agent or the command line.
- [](LICENSE)
- [](https://modelcontextprotocol.io)
- [](skills/sound-forge/SKILL.md)
- [](https://github.com/comfyanonymous/ComfyUI)
- > "make a heavy wooden door slam for the crypt level" → your agent generates four takes,
- > keeps the one that best matches the prompt, trims it, makes it mono, normalises it and
- > checks it - about two seconds later.
- Sister project of [Asset Forge](https://github.com/Noamcelermajer/comfyui-text-to-3d-mcp)
- (text to 3D models); both follow the same profile / preset / skill design.
- **Why Sound Forge**
- - **No credits, no cloud.** Everything runs on your own NVIDIA card, as often as you like.
- - **Game-ready, not just generated.** Every file is trimmed, looped when it should loop
- (equal-power crossfade, cut to whole bars for music), converted to mono or stereo,
- normalised to a loudness target with a peak cap, and checked for silence, clipping,
- gaps and DC offset.
- - **Best of N.** Several takes per call, ranked by how well they match the prompt (LAION
- CLAP); the others are kept so you can swap.
- - **Characters keep their voice.** A voice is designed once from a description, saved,
- and cloned for every later line.
- - **Fast / standard / max presets** and **hardware tiers** (8 / 12 / 16 / 24 GB) chosen
- automatically.
- - **Consistent palette.** Style presets (`retro_8bit`, `cartoon`, `realistic`,
- `dark_fantasy`, `sci_fi`) or a per-game `soundforge.json`.
- - **Swap models without code.** Each kind of sound is a small JSON profile around a
- ComfyUI graph.
- ## What it makes
- | Profile | For | Model | Output |
- | :--- | :--- | :--- | :--- |
- | `sfx` | impacts, footsteps, weapons, spells, pickups, UI | Stable Audio 3 Small-SFX | mono, trimmed, -16 LUFS |
- | `ambience` | forest, rain, cave, tavern, engine room | Stable Audio 3 Small-SFX | stereo, seamless loop, -24 LUFS |
- | `stinger` | level up, victory, game over, item found | Stable Audio 3 Small-Music | stereo, -16 LUFS |
- | `music_loop` | exploration, combat, town, menu | ACE-Step 1.5 (turbo) | stereo, loop cut to whole bars, -18 LUFS |
- | `music_track` | title theme, credits, cutscene | ACE-Step 1.5 (turbo) | stereo, fades out, -18 LUFS |
- | `voice` | dialogue, barks, narration | Qwen3-TTS 1.7B (design + clone) | mono, trimmed, -18 LUFS |
- Presets: `fast` (one take, fewer steps), `standard` (default), `max` (Stable Audio 3
- Medium or more takes). Short hits come out quieter than -16 LUFS: the peak is capped at
- -1 dBFS rather than clipped.
- ## Measured
- RTX 5080 16 GB, standard preset, seed 1234, models already loaded (the first call of a
- session adds 10-20 s while weights load). VRAM is the whole GPU, including what ComfyUI
- already had cached.
- | Job | Result | Time | GPU memory |
- | :--- | :--- | ---: | ---: |
- | `sfx` door slam, 4 takes | 0.99 s, -19.8 LUFS | 20 s cold / ~1-2 s warm | 6.6 GB |
- | `sfx` coins in a pouch, 4 takes | 1.08 s, -18.8 LUFS | 1.4 s | 5.7 GB |
- | `sfx` fire spell, 4 takes | 1.48 s, -18.3 LUFS | 1.0 s | 5.6 GB |
- | `ambience` cave, 3 takes | 28.1 s seamless loop, -24.8 LUFS | 3.3 s | 5.7 GB |
- | `stinger` level-up, 4 takes | 4.1 s, -16.0 LUFS | 1.7 s | 6.0 GB |
- | `voice` design a character voice | saved reference clip | ~20 s | 11.9 GB (first line only) |
- | `voice` line 1 / line 2 (cloned) | 3.9 s / 4.3 s, -18.0 LUFS | 9.8 s / 9.5 s | 8.5 GB |
- | `music_loop` 60 s | *being measured* | | |
- Every job passed its checks. You cannot judge audio from numbers: listen to the results
- (`tests/smoke_test.py` writes them to `tests/results/`).
- ## Requirements
- - NVIDIA GPU with 8 GB of VRAM or more (16 GB tested), Windows or Linux
- - ComfyUI 0.45 or newer (portable build is fine) and git
- - About 35 GB of free disk for the default models (26 GB download), 60 GB with `--all`
- | VRAM | What you get |
- | :--- | :--- |
- | 8 GB | Everything at `standard`. Music makes one take at a time; the voice model unloads after each line. |
- | 12 GB | Everything, including the `max` presets with Stable Audio 3 Medium. |
- | 16 GB | Everything (reference machine). |
- | 24 GB+ | Room for the optional ACE-Step XL music model and its 4B planner. |
- Only the 16 GB row has been measured.
- ## Install
- The easiest way: paste [SETUP_PROMPT.md](SETUP_PROMPT.md) into Claude Code or Codex. By hand:
- ```
- git clone https://github.com/Noamcelermajer/comfyui-text-to-audio-mcp SoundForge
- cd SoundForge
- python forge.py setup # node pack + models for this GPU (--dry-run to preview, --all for the optional ones)
- python forge.py doctor # every profile should be "ready": true
- python forge.py connect # registers the MCP server and the skill for Claude and Codex
- ```
- Sound Forge finds ComfyUI next to it or through `SOUNDFORGE_COMFY`, starts it when needed,
- and hands over to ComfyUI's Python for the audio libraries. Full unattended instructions
- for agents: [AGENTS.md](AGENTS.md).
- ## Use
- From an agent: just ask ("make an ambience loop for a rainy harbour at night"). The skill
- tells it which profile, length and prompt style to use. The MCP server has 7 tools:
- `list_profiles`, `generate_sound`, `design_voice`, `get_job`, `check_audio`, `models`,
- `forge_status`.
- From the command line (every command prints one JSON object):
- ```
- python forge.py sound door_slam "Heavy wooden door slamming shut in a stone hall, deep thud with a short echo" -p sfx --set duration=3
- python forge.py sound cave "Damp cave interior, slow water drips, low wind rumble" -p ambience --project ../MyGame
- python forge.py sound explore "Calm fantasy exploration, acoustic guitar, soft strings, flute" -p music_loop --set bpm=90 --set keyscale="D major"
- python forge.py voice grave_keeper "An old man with a deep gravelly voice, slow and weary"
- python forge.py sound keeper_01 "Who goes there? The graveyard is closed after dark." -p voice --voice grave_keeper --set emotion="suspicious"
- python forge.py check some_file.wav
- python forge.py batch jobs.json
- ```
- Exit code 0 = ok, 2 = written but failed its check (see `problems`), 1 = error.
- ## Customising
- - **Per game:** `python forge.py init <game folder>` writes a `soundforge.json` with the
- output folder, voices folder, style and per-profile settings.
- - **Per call:** `--set key=value` / `params` (`duration`, `candidates`, `loudness`,
- `format=ogg`, `sample_rate=48000`, `loop`, `bpm`, ...). `forge.py profiles` explains
- every parameter.
- - **New profile:** a JSON file in `profiles/user/` or the game's `soundforge_profiles/`
- that `extends` an existing one.
- - **New model:** add it to `models/catalog.json` and write a profile around a ComfyUI API
- graph in `graphs/`.
- Settings are layered: profile, hardware tier, project, preset, style, call; the tier's
- limits are applied last so a call cannot run an 8 GB card out of memory.
- ## Licences
- Sound Forge's code is MIT. The models are not: Stable Audio 3 uses the **Stability AI
- Community License** (free for individuals and organisations under $1M annual revenue; you
- own the outputs), ACE-Step 1.5 is MIT with commercial use of outputs allowed, Qwen3-TTS
- and CLAP are Apache 2.0. Details in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
- Only clone voices you have the right to use.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.