earshot
Ears for your coding agent: earshot describes a sound in words, tells whether it matches what you asked for, finds the right moment in a long recording and checks that a sound effect lands on the picture cut. It also makes sounds in code, from a bank of 252 effects that render with no model and no network (whooshes, hits, clicks, notes). Runs locally on the CPU, as a command line tool, in Python or as an MCP server for Claude Code. MIT licence, version 0.4.1 of November 6, 2026, about 4.6k stars on November 12, 2026.
Strengths
- Measures how sound lines up with picture: the timeline command lists every sound event and its sync with the video cuts
- 252 effects rendered from code in seconds, with no model and no network, with adjustable length, peak point and pitch
- Ranks candidate files against a description: handy to pick the right sound among several you downloaded
- Everything runs locally on the CPU, no graphics card, and plugs into Claude Code as an MCP server
Limitations
- Sounds made in code remain synthetic: the published blind test rests on one listener and fifteen trials, and the author himself says it is not a proof
- It judges neither taste, nor emotion, nor the licence of a sound: a file found elsewhere is still yours to check
- Technical setup: uv and FFmpeg required, two models of about 1.1 GB to download, and each analysis call on the command line reloads the models in 20 to 40 seconds (measure, sfx and kit load none)
- Repo created in October 2026, already at version 0.4: commands change fast, read the changelog before updating
Best for
- Creators who have an agent edit their videos (HyperFrames, Remotion) and want sound effects synced to the cut