中文版

earshot（GitHub：https://github.com/Leonxlnx/earshot ，2026-10-15 创建，今晨抓取时已 4,078★）是给编码 Agent 装“耳朵”的开源工具：Agent 现在能写渲染影片、产品演示或游戏的代码，却听不见自己做出来的东西——分不清高级的按键声和廉价的按键声，听不出音效比画面慢了 80 毫秒。earshot 把声音变成 Agent 能推理的证据：一段文件听起来像什么、与描述贴合到什么程度、长录音里最精彩的时刻、是真实录音还是代码合成的、一整部影片的逐秒听觉记录、带标注的声图，以及对照专业发布影片的评分卡。本机 CPU 运行，提供 CLI、Python API 与 MCP 服务器，MIT 开源。

命令覆盖很全：describe 描述声音、match 判断贴合度、find 定位长录音里的最佳时刻、realness 鉴别真实与合成、timeline 生成听觉记录、sheet 输出带标注的声图、critique 给发布影片打分并指出带时间戳的修复项。还能反向做声音：forge 用代码写音效、边听边迭代改进，fit 把一段录音翻译成可调参数的合成配方；自带 252 个代码音效的音库，sfx/kit 可直接按画面时刻渲染整部影片的音效。

为什么值得了解

• 对用 Agent 做影片、演示或游戏的人，这是补上最后一块感官拼图：画面能看、代码能跑之后，声音终于也能被 Agent 自己检查和迭代
• forge + fit 的玩法值得玩声音的人参考：音效是代码配方而非录音文件，可以像改代码一样调参、变体、复用
• 但每次 CLI 调用要加载模型（约 20~40 秒），少量调用没问题，多了建议走 MCP 服务器或 Python 脚本常驻；可选的音频大模型点评只在你明确同意并自备 Key 时才联网

English Version

earshot (GitHub: https://github.com/Leonxlnx/earshot , created 2026-10-15, 4,078 stars at this morning’s fetch) is an open-source tool that gives coding agents ears: agents can now write the code that renders a film, a product demo or a game, but they cannot hear what they made — they cannot tell a premium key press from a cheap one, or hear that a hit lands 80 ms after the cut. earshot turns sound into evidence an agent can reason about: what a file sounds like in words, how well it fits a description, the best moment inside a long recording, whether it is a real recording or code-synthesised, a second-by-second listening transcript of a whole film, an annotated picture of it, and a scorecard against professional launch films. It runs locally on CPU, as a CLI, a Python API and an MCP server, MIT licensed.

The command set is broad: describe narrates a sound, match judges how well a description fits, find locates the best moments in a long recording, realness tells real from synthesised, timeline writes the listening transcript, sheet outputs an annotated sound picture, and critique scores a launch video with timestamped fixes. It also works in reverse: forge writes a sound in code and improves it by listening, fit translates a recording into a synthesis recipe with knobs and morphs, and a shipped bank of 252 code-defined sound effects can be rendered against picture moments with sfx and kit.

Why it’s worth knowing

• For anyone using agents to make films, demos or games, this fills in the last missing sense: once the picture looks right and the code runs, the sound can finally be checked and iterated on by the agent itself
• The forge plus fit approach is worth exploring for anyone working with sound: effects are code recipes rather than recorded files, so they can be tweaked, varied and reused like code
• Each CLI call loads the models first (about 20–40 seconds), though — fine for a few calls, but use the MCP server or a Python script for heavier use; the optional audio-LLM critique only contacts the network with your own key and explicit consent