earshot

About

Ears for coding agents, CPU only and open source

Agents can now write the code that renders a product demo, a game or a launch film, but they cannot hear the result. earshot fills that gap. `describe` labels a sound, `match` and `rank` score files against a text description, `find` locates the best trims inside a long recording, `realness` tells a real recording from a code-synthesized one, `timeline` produces a listening transcript with sync to picture cuts, and `critique` scores a launch video's sound against a profile of professional films with timestamped fixes. `forge` lets an agent build a sound in code and improve it by listening, and `fit` translates a recording into code it can vary. It ships an MCP server for Claude Code and other clients (models stay warm between calls) plus two agent skills. Requires Python 3.10+ and ffmpeg; no GPU. Repo: https://github.com/Leonxlnx/earshot.

Features

• Describe, match and rank sounds against text descriptions
• Find the best moments inside long recordings
• Listening transcript and annotated PNG sheet for video sound
• Launch-video sound critique with timestamped fixes
• Make sounds in code and refine them by listening
• Real vs synthesized detection
• MCP server plus two agent skills, CPU only

Install