earshot: open-source ears for AI coding agents
AI agents can write the code that renders a film, but they can't hear it. earshot turns sound into words and numbers an agent can reason about, and it runs on a plain CPU.
Suresh26 Oct 20262 min readAI coding agents can now write the code behind a product video, a motion-graphics reel or a small game. What they still can't do is hear what they made. They can't tell a premium key press from a cheap one, and they won't notice that a hit lands 80 ms after the cut.
earshot is an open-source project that gives agents a pair of ears.
What it does
earshot turns audio into evidence an agent can read: a description in words, scores, timings and measurements. It ships as a command-line tool, a Python API and an MCP server, so agents such as Claude Code can call it directly.
- describe: what a sound is most like, with calibrated tags and the closest alternatives
- match: how well a file fits a description, with a verdict from strong to no match
- find: the best moments inside a long recording, as trim points
- realness: whether a sound looks like a real recording or something synthesized in code
- timeline and sheet: a second-by-second listening transcript of a whole film, plus an annotated picture the agent can look at
- critique: a scorecard for launch videos against a profile of seven professional launch films
forge: sound made in code
The most fun part is earshot forge. The agent writes a sound in code, renders it, listens to the result and keeps improving it. Every sound in earshot's own launch film was made this way, with no recordings and no music track. Newer versions also ship a bank of ready-made code sounds (whooshes, hits, clicks, notes) that render in seconds.
Free, and CPU only
- Licence: MIT for the code. The two models it downloads (CLAP and an AudioSet tagger) keep their own open licences.
- Hardware: no GPU needed. It needs Python 3.10+ and ffmpeg.
- Offline: after a one-time download of about 1.1 GB of models, it runs without a network.
Honest limits
The README is upfront about what earshot can't do. It doesn't judge taste or emotion, doesn't understand what is being said in speech, and its models don't hear anything above 14 kHz. Tags for events buried in a busy mix are hints, not facts.
If you make videos with code, as we do for SureShot Tech, this is a tool worth trying.
