earshot

Agent skill by Leonxlnx · for Claude Code, Codex, Cursor and OpenCode

A CPU-only CLI, Python API and MCP server that lets coding agents hear, rank, trim, verify and forge game and film sound effects, with a bank of 252 code-made sounds and sync checks against picture cuts.

  • Engine: Engine-agnostic
  • 4 agents
  • License: MIT
  • Free
  • 5.3k stars

Install

uv tool install "earshot[mcp] @ git+https://github.com/Leonxlnx/earshot"

Run it in a terminal, then start a new agent session so the skill loads. Read the repo's README and scripts before you install: skills run with your agent's permissions.

What it does

earshot gives AI coding agents ears. Agents that can write game or launch-film code still cannot hear whether a hit lands late, a key press sounds cheap, or the music swells in the wrong place. earshot turns audio into evidence: calibrated descriptions, match scores against a phrase, best trims inside a long recording, real-vs-synthesized checks, listening transcripts, annotated sheets and launch-film scorecards.

As of 2026-11-06 it ships as open-source MIT software that runs locally on CPU. Install with uv, then optionally expose the same tools over MCP for Claude Code. Two Agent Skills live in the repo: earshot (how to listen and which command answers which question) and real-sound (a process for premium soundtracks from licensed recordings). A forged bank of 252 code-made effects covers UI pops, hits, whooshes, notes and launch-film style sounds, renderable without models or network calls.

Forge and fit loops let an agent make or translate sounds in code and improve them by listening. Timeline and kit tools place effects against picture cuts and music beds. First use downloads two pinned open models (~1.1 GB) for description and tagging; after that it can stay offline. Needs uv, ffmpeg/ffprobe, and enough disk for the model cache.

Features

  • Describe, match, rank and trim sounds with plain-language verdicts
  • MCP server so models stay warm across agent calls
  • 252 code-made SFX bank plus forge and fit loops
  • Timeline and kit tools for sync to picture cuts
  • Two shipped Agent Skills: earshot and real-sound
  • CPU-only; optional offline mode after model fetch

Limits

  • Needs uv plus ffmpeg/ffprobe on PATH
  • First model use downloads ~1.1 GB; cold CLI calls are slow vs MCP
  • Optional Gemini critique needs your own API key and explicit consent

Guides that cover earshot

Browse all 84 agent skills