Software
Sound Effects and Music Kit
From this video: Free Sound Effects and Music With Claude
Sound effects and background music for your videos, with a voiceover or without one. It runs on your own Windows PC with free models and a free pack of real recorded sounds. Nothing is locked.
This is everything from the video, in one zip.
Free, and no email needed to download. You don't type commands: Claude does the setup, and you click a few links and say yes.


This needs a computer. On a phone? Email yourself the link for later.
Your download has started. That's everything you need. Curious what Sound Effects and Music Kit Pro adds anyway? It's all side by side.
Compare the two- No email and no card needed
- Zip, 3.5 MB
What you get
- The guided setup: open the folder in Claude's Code tab and say "set up my sound". It asks which parts you want, then downloads them
- The step list Claude follows, and the sound rules, with a place for your own
- The sound tool for videos with a voiceover
- The sound tool for videos without one
- A search for the free sound pack
- The checks: loudness, peaks, hiss, and a word check for videos with a voiceover
- The connection that gives Claude its sound tools
- What to say to Claude, and the exact messages to send if you would rather not use the zip
- A short starter clip of our own to try it on
- 3 example plans
What each one has
| Question | Sound Effects and Music Kit | Sound Effects and Music Kit Pro |
|---|---|---|
| Can I add sound effects and music to my video the way the video shows? | Yes, all of it | Yes, the same files |
| Do I need the big model downloads to get music? | Yes, or music you bring yourself | No: 48 pieces are in the music zips |
| Does it work on a PC that can't run the models? | Yes, with the sound pack and your own music | Yes, and the music and effects are already there |
| Do I get ready-made effects? | No: it makes them, or finds them in the free sound pack you download | Yes, 62 |
| Can I see the real plans behind the video? | 3 example plans for the starter clip | All 11 parts of the video, and the clip four ways |
| Can I check my sound against a video I like? | No | Yes |
| Can I check I'm not using the same music as last time? | No | Yes |
| How big is the download? | 3.5 MB | 374 MB, in three zips |
| Do I need to give an email? | No | Yes, checkout needs one for the receipt |
How it runs
- Open the kit's folder in Claude's Code tab and say: set up my sound. Claude asks which parts you want (effects, quiet music, full songs, or only the sound pack) and downloads them.
- Put your video in the my-videos folder and say: make the sound for this. Claude shows you the plan first, so you can fix it before it starts.
- Effects come from the free sound pack first and the effects model second. Background music comes from the quiet music model.
- You get the mix and a short report with the spots to listen to.
The catch
- Windows only. It needs the Claude desktop app with the Code tab, and Git for Windows.
- About 30 GB of free disk for everything, less for fewer parts. The setup took about an hour and a half in our run, most of it downloads.
- It works without a graphics card, just slower. On a fast PC with the card switched off, a 4 second effect took 13 to 21 seconds, and 60 seconds of quiet music took about 2 minutes. Not tested on a real AMD or Intel card, or on an old laptop.
- The models, the engine and the sound pack are separate downloads under their own licences. The kit ships none of them. You click the sound pack's five download links yourself, because its site checks for a human.
- Stable Audio's licence is free under 1 million US dollars a year, and asks for a free sign-up if your channel makes money.
- Nobody has listened to what the kit made in its tests. Its checks are measurements, so your ears decide.
What is in the zip
- START-HERE.md: what you need, the disk space table and the three steps to start
- CLAUDE.md and rules/sound-rules.md: the step list Claude follows and 17 sound rules, with a place for your own
- messages: what to say to Claude, and the no-download route (the exact messages to send by hand)
- setup and tools: the guided setup, the sound tool for videos with a voiceover and without one, the sound pack search, the checks, and the connection that gives Claude 11 sound tools
- example: a 5 second starter clip of our own and three example plans
- LICENSE.txt and credits-and-licences.md
After you download
- Unzip the kit on a drive with about 30 GB free. Keep the path short: a folder called sound-kit right on that drive is best.
- Open that folder in Claude's Code tab.
- Say: set up my sound
- Then put your video in the my-videos folder and say: make the sound for this
Rather not download a zip?
Send these messages to Claude instead, one at a time, in an empty folder open in Claude's Code tab. It is the same text as messages/no-download-route.txt in the kit.
NO-DOWNLOAD ROUTE
Don't want to download the kit? Make an empty folder (for example a folder called sound right on a drive), open it in Claude's Code tab, and send these
messages one at a time. Wait for Claude to finish each one. This builds the same setup the kit has. The kit is tested;
what Claude writes from these messages is new code each time, so check its first results more carefully.
--- MESSAGE 1 ---
Set up free local sound for my videos in this folder, on Windows, and keep everything inside this folder. Ask me one
question at a time. First check that Git is installed (if not, tell me to install Git for Windows and stop). Get uv into
a bin folder here if I don't have it, and make a Python 3.12 environment in .venv with uv, with uv's cache and Python
inside this folder too. Install numpy, soundfile, httpx, psutil and mcp. Check for ffmpeg; if it's missing, get the
"release essentials" zip from gyan.dev into bin. Then tell me my graphics card and its dedicated memory, and my free disk.
--- MESSAGE 2 ---
Ask me which parts I want: effects, quiet music, full songs, or only the sound pack. Then download what I pick, from these
official places only, with a download script that splits the file into pieces, asks for byte ranges, appends a piece only
when the server answers 206, picks up where it stopped if the connection drops, and checks the final size:
- the audio.cpp engine v0.9.3 from github.com/0xShug0/audio.cpp releases: with an NVIDIA card
audio-v0.9.3-bin-windows-x64-cuda12.4.zip (465913709 bytes) plus audio-v0.9.3-cudart-windows-x64-cuda12.4.zip
(607273675 bytes, its DLLs go next to the program); with another card or none
audio-v0.9.3-bin-windows-x64-cpu.zip (30547092 bytes). (A vulkan build exists too, but it crashed on the first
effect when we tested it.)
- models from huggingface.co/audio-cpp/audio.cpp-gguf (resolve/main/...):
effects Stable-Audio-3-Small-SFX-GGUF/stable-audio-3-small-sfx-f16.gguf (2358472992 bytes),
quiet music Stable-Audio-3-Medium-GGUF/stable-audio-3-medium-q8_0.gguf (3582601408 bytes),
full songs ACE-Step1.5-GGUF/turbo/ace-step-1.5-turbo-q8_0.gguf (6185468096 bytes).
Put each model in its own folder (models/sfx, models/bed, models/music) and write the engine's server.json: host
127.0.0.1, port 8030, backend cuda or cpu to match the build, lazy_load true, max_loaded_models 1, and one entry per
model: ids sfx and bed with family stable_audio, id music with family ace_step, task gen, mode offline.
--- MESSAGE 3 ---
I'll download the free Sonniss GDC 2026 sound pack myself from https://gdc.sonniss.com (five zips; the site checks for a
human, so don't try to get past that). They are in: <the folder they landed in>. Unzip them into sound-pack/sounds and
build a search index in SQLite (full-text search over folder and file names split into words, CamelCase split, library
codes like WHSH and IMPT turned into whoosh and impact hit, a few synonyms; files with several takes cut at silent gaps
of 0.25 s). The search must need every word to match. Leave my zips where they are. Never copy these sounds anywhere else:
the licence allows them in my videos but not sharing the files or training AI on them.
--- MESSAGE 4 ---
Write a sound studio in Python that talks to the engine (POST /v1/tasks/run with {"model", "request": {"text",
"options"}}), and follow these rules every time:
- effects: the sound pack first, the effects model second; search with two or three plain words; skip pack names with
sci fi, magic, christmas, cartoon, toon, game, fantasy, laser, alien, comedy, funny, horror, monster.
- never ask the effects model for under four seconds; trim afterwards; measure every made sound for hiss (spectral
flatness of 0.35 or more, or 0.3 or more with the centre above 7 kHz, is hiss); three tries, then drop the sound.
- quiet music: three takes, keep the cleanest (brightest, fewest clipped samples); full songs: two takes.
- end every music description with: no ukulele, no whistling, no glockenspiel, no claps, no risers, no braams, no
choir, no ticking clock. One small band for the whole video; scenes only get calmer or busier.
- engine audio can clip: if any sample hits full scale, ffmpeg volume=-3dB,adeclip, then peak to -1 dBFS.
- with a voiceover: voice brought in at -16 LUFS, music at -20 dB and dipping 10 dB under the voice (sidechain), effects
-18 to -26 dB and 6 dB lower where they land on a word; final mix at -14 LUFS, true peak -1 dB.
- with no voiceover: a plan of moments seen in the picture (pull frames out of the video with ffmpeg into contact
sheets stamped with their times, and read them); effects and music tied to moment ids; fx, music and mix files exactly
as long as the video.
- sounds only on something you can see happen or hear said. A change I ask for changes only that sound; everything else
is reused, so the music stays the same piece.
- check every mix (loudness within 1 LU, true peak) and keep the last good mix if a check fails.
Write these rules into rules/sound-rules.md, with a "Your rules" section at the bottom for my notes.
--- MESSAGE 5 ---
I make videos with a voiceover: install openai-whisper into .venv, with the base.en model kept in a checker-models folder
here. Use it to find where each line of my script is said, and after every mix turn the mix back into text and compare
it with the bare voice: a word heard in the voice but missing in the mix fails the mix. (Skip this message if you have
no voiceovers.)
--- MESSAGE 6 ---
Write a small MCP sound server (Python, FastMCP) with tools to find a sound, prepare a pack sound, get an effect by the
rules, make an effect, make quiet music, make a song, extend a loop, mix, run a whole sound plan, measure hiss, and show
the status. Back up my Claude user settings file (.claude.json in my user folder) first, then add exactly one entry
under "mcpServers" that starts it with this folder's .venv Python. Show me the backup. Then test everything: start the
engine, make one four second test effect, measure it for hiss, stop the engine, and tell me how much disk it all takes.
--- AFTER THAT ---
Say "make the sound for this" with your video in the folder. Ask Claude to show you the plan first.Credits
- Engine: audio.cpp by ShugoAI LLC (Apache 2.0).
- Effects and quiet music: Stable Audio 3 Small-SFX and Stable Audio 3 Medium by Stability AI, under the Stability AI Community License. Powered by Stability AI.
- Full songs: ACE-Step 1.5 Turbo by ACE Studio and StepFun (MIT).
- Word checker: Whisper base.en by OpenAI (MIT).
- Recorded sounds: the Sonniss GDC 2026 Game Audio Bundle, under its own licence (sonniss.com/gdc-bundle-license). Not in the kit.
What it does not include
- Sonniss sounds are not included. The free pack is a separate download under its own licence, and you click its download links yourself.
- No engine and no models inside. The setup downloads them from their makers, under their own licences.
What you need
- A Windows PC. Windows only.
- The Claude desktop app with the Code tab. It is not included. Web chat does not work.
- Git for Windows. Claude's app needs it on Windows.
- About 30 GB of free disk for everything. Less if you pick fewer parts.
- Your video. With a voiceover: the voice file and its script as text.
- A graphics card helps but is not needed. Without one it is slower.
Details
- Device
- Needs a computer
- Tested on
- Windows
- File
- Zip, 3.5 MB
- Version
- 1.0.0, 13 November 2026
Updates are free. The download always gives you the newest version.
Licence and credits
Our files are under the MIT licence (LICENSE.txt): free to use and change. The parts the setup downloads keep their own licences, listed in credits-and-licences.md. Nothing here limits those.
Not affiliated with Anthropic. Claude is a trademark of Anthropic, PBC.