WhoSound-curious developer who wants a custom click, whoosh or key press made entirely in code
The problemStock effects never quite fit, and an agent cannot judge a synthesized sound without hearing it.
The outcomeA folder of rendered WAVs, a report in the critic's words, spectrogram comparisons and an audition file.
Steps
Run earshot forge --list to see recipes like click, impact, keypress and whoosh.
Run earshot forge keypress --refs path/to/real/keypresses/ --budget 144 --variants 4 --dir made/keypress with your own reference folder.
Open report.md and compare.png in the output folder and listen to audition.wav.
Re-render a pick with earshot forge --render keypress --params made/keypress/report.json --seeds 1 2 3 --dir made/again.
Prompt for your agent
Use earshot forge to make four variants of a <sound type> against my reference folder <refs dir>, then read report.md and tell me which pick is closest to the references. Verify by listening to audition.wav, not just the scores.
Caveat. The search optimizes earshot's own models and can fool them (the README notes this), so check spectrograms and listen; the v1 blind test was a coin flip on one listener and 15 trials.
Rank candidate files against a text description and trim the best moment from a long recording.
Intermediate
45 minutes
Any agent
Seen in the wild
Evidence (similar to what people reported): X, Oct 2026
WhoGame or video maker who downloaded several candidate sound effects and needs to pick one
The problemChoosing among many similar files by ear is slow, and long recordings hide the one good hit.
The outcomeA ranked list of files for your description, plus start and end seconds for the best moment in a long file.
Steps
Run earshot rank QUERY PATHS... with a text description and your candidate files.
Run earshot find keys.ogg "a single mechanical key press" --top 4 on a long file to get trim points.
Check a top pick with earshot match key.ogg "a deep thocky mechanical keyboard press" --alts "a clicky blue switch keyboard" "a mouse click".
Trim the file at the returned start and end seconds.
Prompt for your agent
Use earshot rank to order the files in <folder> by <description>, then run earshot find on the top file to give me trim start and end seconds. Verify the top pick with earshot match against two alternative descriptions.
Caveat. Matching quality depends on the built-in vocabulary of 322 descriptions, so unusual sounds may rank loosely.
WhoBuilder with a folder of unlabeled sound files who needs to know what each one actually sounds like
The problemAn agent cannot open a .wav or .ogg and tell what it is, how long it rings or how bright it is.
The outcomeA short text reading with calibrated tags, closest alternative readings and numbers like peak level and brightness.
Steps
Install earshot and run earshot doctor.
Run earshot describe enter.ogg (swap in your file).
Add --json if the agent needs the full structured result, or --out f.json to save it.
Use the description to label or sort the file in your project.
Prompt for your agent
Use earshot to describe every audio file in <folder> and write a one-line label per file from the tags. Verify one label by listening or by running `earshot match` on it.
Caveat. Tags are model readings with a confidence score (z), so treat low-confidence results as hints, not facts.
WhoFounder or motion designer finishing a launch video and unsure if the mix is up to standard
The problemThere is no objective way to know if the soundtrack is too quiet, too thin or late compared to professional launch films.
The outcomeA verdict with scored metrics and a numbered list of timestamped fixes, worst first.
Steps
Export your film as a video file with audio.
Run earshot critique film.mp4 on it.
Read the VERDICT and the FIX FIRST list.
Apply the top fixes in your audio code or editor, then run the critique again.
Prompt for your agent
Run earshot critique on <film.mp4> and summarize the verdict and the top three timestamped fixes. Verify by re-running critique after the changes and comparing how many metrics are in range.
Caveat. The scorecard compares against a profile of 7 professional launch films, so it suits launch and promo styles, not every genre.
WhoDeveloper who has one good sound and wants endless variations or a blend of two sounds
The problemA fixed audio file cannot be resized, retuned or varied, so repeated use sounds identical.
The outcomeA recipe module that re-renders the sound with new takes, readable knobs and morphs between two recordings.
Steps
Run earshot fit click.wav -o made/click_fit.py --report on your recording.
Run the module with python made/click_fit.py 3 take3.wav to render a new take.
Blend two sounds with earshot fit --morph a.wav b.wav --amount 0.5 -o made/ab_morph.py.
Explore variations with earshot fit --forge made/click_fit.py --budget 96 --dir made/click_variations.
Prompt for your agent
Use earshot fit to convert <recording.wav> into a recipe module, then render three new takes and describe how they differ. Verify the fit by comparing the side-by-side spectrogram in the report.
Caveat. The fit stores the structure of the sound, not its samples, so very complex or noisy recordings may reproduce only approximately.
WhoDeveloper making a code-rendered film or app demo who needs whooshes, hits and clicks fast, with no recordings
The problemHunting for and syncing sound effects by hand is slow, and the agent cannot hear what it picked.
The outcomeA kit.json plus rendered takes for every role, each with measured onset, peak and end times.
Steps
Run earshot sfx list to see the bank of 201 sounds.
Run earshot kit init to start a kit and choose a sound and level per role in kit.json.
Run earshot kit render to render the cached takes, then earshot kit check to read the measured sync points.
Run earshot sfx render ID -o x.wav to aim any single sound at a moment with --dur, --apex or --pitch.
Prompt for your agent
Use earshot kit to build a sound kit for <film description> with one bank sound per role, render it, and report the measured onset and peak of each. Verify with `earshot kit check` that no role is missing.
Caveat. The sfx and kit commands need no models and no network, but the v0.3 bank is new and the README says no blind-test score for v2 sounds is published yet.
WhoDeveloper using Claude Code or another MCP-capable agent on audio-heavy projects like films and games
The problemCalling the CLI repeatedly reloads models for 20 to 40 seconds each time, which makes listening loops painfully slow.
The outcomeThe agent calls describe, match, timeline and critique as tools with models kept loaded.
Steps
Install earshot and run earshot doctor --load to confirm models load.
Start the server with earshot mcp.
Register it in your agent's MCP settings (inferred, follows your agent's usual MCP config).
Ask the agent to listen to a file and describe it to confirm the tools respond.
Prompt for your agent
Connect to the earshot MCP server and use its tools to describe <audio file> and critique <video file>. Verify the tool list shows the listening tools before relying on them.
Caveat. The README says the MCP server works for Claude Code and other agents but does not give a config snippet, so the registration step depends on your agent.
WhoDeveloper whose coding agent renders a product demo video and wants the audio to line up with the cuts
The problemThe agent can render a film in code but cannot hear whether a hit lands 80 ms after the cut.
The outcomeA timeline of sound events, loudness and sync results against the cuts, so the agent can adjust timings.
Steps
Install earshot and run earshot doctor to confirm models and data load.
Run earshot timeline FILE --video V on the rendered film (replace FILE and V with your paths).
Read the events and sync section, then have the agent shift the offending cues.
Re-render and run the timeline again to confirm the offsets shrank.
Prompt for your agent
Use earshot to run a timeline on <film.mp4> with its audio and video, then list every sound event that misses a picture cut and by how many milliseconds. Verify by re-running the timeline after your fix.
Caveat. Each CLI call reloads the models (about 20 to 40 seconds), so use the MCP server or one Python script for repeated checks.
Score files for realness and compare them to reference recordings or each other.
Intermediate
30 minutes
From the README
Any agent
WhoAudio-minded builder checking whether a sound in a project reads as a real recording or an obvious synthesis
The problemIt is hard to tell, without careful listening, whether a sound will pass as real or give itself away.
The outcomeA real versus synthesized verdict with evidence and per-window scores, plus a plain-words A/B of two sounds.
Steps
Run earshot realness FILE... on the files you want to check.
Run earshot ab A B to compare two sounds or two trims in words.
Run earshot compare FILE --refs DIR to see the nearest reference recordings.
Use the findings to decide which sound to rework.
Prompt for your agent
Use earshot realness on <files> and earshot ab on the two closest candidates, then tell me which one reads as most real and why. Verify with a listening check, since the realness model can be fooled.
Caveat. The README warns the realness model is easy to fool (a modal resonator plus a room and noise floor reads as real), so do not treat the verdict as proof.
Who it is forBuilt for people who use coding agents to make films, product demos and games in code, and need the agent to actually hear and judge the sound it produces.
Why it is hot+711 stars today, 1.4k total, launched 4 days ago
DEDUPED