Skip to main content
Glama

NoiseSmith

Describe a sound in plain words inside Claude, and get an audio file saved in your project.

"an old creaky door" · "wind, from a slight breeze to a hurricane" · "a heavy iron gate slamming" · "light drizzle"

Status: early development (Phase 1). NoiseSmith can make wind, whooshes, door and floor creaks, impacts, rain and game blips today, and loop, layer and export them. Animal sounds (a barking dog) are planned for a later AI-powered upgrade. See PROJECT_PLAN.md.

Free and open source (MIT). Everything the built-in engine makes is CC0: use it anywhere, including commercial projects, no credit needed. Some sounds are synthesized from scratch (whooshes, impacts, game blips). The natural ones are made from real recordings that are reshaped, because synthesizing them never sounded natural: the creaks (see credits) and the wind, rain and thunder, recorded by Joseph Sardin (see credits). All are CC0.

What it can make

Recipe

Examples

Settings

wind

breeze, gale, hurricane (steady, gusty or howling)

strength

thunder

distant rumble to a close crack

how close, storm size

whoosh

swish, sword swing, passing car, jet fly-by

power, object size

door_creak

cupboard, old door, heavy gate, long horror-movie creak

how worn, door size

floor_creak

creaking floorboards, stairs, old wood

how worn, size

impact

wood knock, metal clang, glass clink, stone thud, boom

how hard, object size

rain

drizzle to downpour

strength

ui_blip

beep, click, coin, jump, error, success, power-up, laser

harshness, pitch

Ask for several versions at once ("give me 5 variations") and pick your favourite. Every sound also gets a small .json file recording exactly how it was made, so it can be re-created.

Related MCP server: spotter

Loops, scenes and files

  • Seamless loops: "a 20 second loop of strong wind". WAV loops carry loop points that game engines read.

  • Scenes: "a storm with gusty wind, heavy rain and a distant boom every 8 seconds" layers several sounds with their own volume, timing and repeats. You can also add custom noise or tone generators for machinery and sci-fi sounds.

  • Formats and loudness: save as WAV, FLAC, OGG or MP3, and match a loudness target: game (-16 LUFS), web (-14), broadcast (-23), or any number.

  • Existing files: convert, trim, resample or loop any WAV, FLAC, OGG or MP3 file.

Realistic animal sounds (optional AI)

Animals and other complicated real-world sounds can't be made well from scratch. For these, NoiseSmith can ask an AI service, ElevenLabs, using your own key. It is optional, off until you set it up, and every sound is billed to your own ElevenLabs account (a few tenths of a cent each on a small plan). Ask for "a very big dog barking" and "a small dog barking" and you get clearly different sounds.

  • Set up: create a key at https://elevenlabs.io/app/api/api-keys (give it a credit limit), then save only the key in ~/.noisesmith/elevenlabs.key, or set the ELEVENLABS_API_KEY environment variable. In the Claude Desktop extension there is a settings field for it. Never paste a key into a chat.

  • Cost safety: an estimate is shown before large requests, anything over about 1000 credits needs your confirmation, and at most 4 variations are made at once.

  • License: the ElevenLabs free plan is non-commercial and needs credit to ElevenLabs; paid plans include a commercial license. Each AI sound's notes file records the prompt, the plan and the license that applies. These sounds are not CC0 like the built-in ones.

Install

You need uv installed. Then, for Claude Code:

claude plugin install noisesmith --marketplace zimba7768/noisesmith

Now just ask Claude: "make the sound of a strong wind howling" or "3 variations of a small wooden door creaking". Files are saved in a sounds folder in your project.

Prefer a plain MCP server (works in Cursor, VS Code and others too)?

claude mcp add noisesmith -- uvx --from git+https://github.com/zimba7768/noisesmith noisesmith

A one-click Claude Desktop extension (.mcpb) can be built with uv run python scripts/build_mcpb.py; downloads will be attached to releases later.

Development

uv sync
uv run pytest

Adding a new sound recipe is meant to be approachable: copy a file in src/noisesmith/recipes/, change the sound, and list it in recipes/__init__.py.

Available Tools

6 tools
create_soundA

Create a sound effect and save it as an audio file (plus a small .json notes file).

Choose the closest recipe (see list_recipes) and translate the user's words into numbers: intensity (0 to 1) is how strong or hard the sound is, size (0 to 1) is how big the source is, style picks a variant. For example "a slight breeze" is recipe wind with intensity 0.1, and "a hurricane" is wind with intensity 1.0. What intensity and size mean depends on the recipe.

variations makes several different versions (1 to 20) so the user can pick one. seed repeats an earlier result exactly. description is the user's original request, saved in the notes file. Files go in ./sounds by default and never overwrite older ones.

loop=True makes a seamless loop of duration_s seconds (default 10), good for ambience such as wind, rain or hums; it is the wrong choice for one-off hits. WAV loops also carry loop points that game engines use. format is wav (default), flac, ogg or mp3. loudness matches a loudness target: "game" (-16 LUFS), "web" (-14), "broadcast" (-23), or a number such as -18. Leave it out to keep the recipe's natural level.

Returns the saved file paths and levels. The sounds are made by NoiseSmith's built-in engine (synthesis, plus reshaped CC0 recordings for the creaks) and are CC0: free to use anywhere, including commercial projects.

Inputs: recipe (required), loop, name, seed, +10 more.

export_soundA

Convert a sound file to another format, level or length, saving a new file.

path is any wav, flac, ogg or mp3 file. format is wav, flac, ogg or mp3. loudness is "game" (-16 LUFS), "web" (-14), "broadcast" (-23) or a number of LUFS; peaks are kept 1 dB below full scale, so a very peaky sound may come out quieter than the target (the result says so). sample_rate resamples (for example 44100). trim_silence removes silence at both ends, and the fades are in seconds. The original file is untouched.

Inputs: path (required), name, format, loudness, +5 more.

list_recipesA

List every sound recipe NoiseSmith can make, with what each setting means.

Call this first if you are unsure which recipe fits, or what intensity and size mean.

Inputs: no named parameters. Returns: result

make_loopA

Turn an existing sound file into a seamless loop, saving a new file.

The last crossfade_s seconds are blended into the start so it repeats with no click. Use trim_s to cut that many seconds off both ends first (to drop a fade-in or fade-out). The result is shorter than the original by the crossfade plus the trims, and the sound should be at least three times the crossfade long. Works best on ambience such as wind, rain or hums. format defaults to the source file's format; WAV loops carry loop points.

Inputs: path (required), name, format, trim_s, +3 more.

mix_soundsA

Layer several sounds into one scene, such as a storm, a busy street or a haunted room.

layers is a list of dictionaries. Each has exactly ONE source: {"recipe": "wind", "intensity": 1.0, "style": "howling"} a NoiseSmith recipe {"file": "sounds/thunder.wav"} an audio file (wav, flac, ogg, mp3) {"generator": {"type": "noise", "color": "brown", "duration_s": 3, ...}} a custom sound A recipe layer also takes that recipe's intensity, size, style, duration_s and seed. Every layer can set: gain_db (volume in decibels, 0 default, -6 half as loud), start_s, length_s, fade_in_s, fade_out_s, speed (below 1 lower and slower), and repeat_every_s with repeat_jitter (0 to 1, uneven spacing), repeat_count and speed_jitter to repeat the sound through the scene. Example, a storm: [{"recipe": "wind", "intensity": 0.9, "style": "gusty", "duration_s": 20, "gain_db": -3}, {"recipe": "rain", "intensity": 0.8, "duration_s": 20, "gain_db": -6}, {"recipe": "impact", "style": "stone", "size": 1.0, "intensity": 0.6, "gain_db": -12, "start_s": 4, "repeat_every_s": 9, "repeat_jitter": 0.6}] A custom generator is {"type": "noise" or "tone", "duration_s", "color" (white/pink/brown), "wave" (sine/triangle/square/saw), "freq_hz" (a number or {"from": hz, "to": hz}), "filters" (a list of lowpass/highpass/bandpass with freq_hz and q) and "envelope" (attack_s, hold_s, decay_s, release_s)}. Generators are for machinery, sci-fi and similar sounds, not animals or voices.

duration_s is the length of the whole scene (up to 120 s). loop, loop_crossfade_s, format and loudness work as in create_sound. If the layers add up too loud, the mix is turned down to stay clear of clipping. Scenes made only from recipes and generators are CC0; a file layer carries whatever license that file has.

Inputs: layers (required), duration_s (required), loop, name, +6 more.

preview_soundB

Play a saved sound through the computer's speakers so the user can hear it.

Inputs: path (required). Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.9/5.0

    Scored across 6 tools

    Disambiguation4/5

    Each tool targets a distinct stage of the sound workflow: recipe discovery, single-sound creation, multi-layer mixing, export/conversion, looping, and playback. There is minor overlap between create_sound's loop option and make_loop, and between mix_sounds and create_sound for complex layered sounds, but the descriptions clarify the different inputs and use cases.

    Naming Consistency5/5

    All tool names follow a consistent snake_case verb_noun pattern: list_recipes, mix_sounds, create_sound, export_sound, make_loop, preview_sound. No mixed conventions or vague verbs.

    Tool Count5/5

    Six tools are well-scoped for a sound-design utility, with each tool covering a clear, non-redundant stage of the workflow. The count avoids both bloat and thinness.

    Completeness4/5

    The set covers the core lifecycle: discovering recipes, creating sounds, mixing scenes, converting/exporting, looping, and previewing. Minor gaps exist around managing saved files—there is no tool to list, delete, or rename generated sounds—but these are workaroundable and do not block primary workflows.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers