leonxlnx/earshot is a local CPU-only tool that gives AI coding agents hearing. Its goal is to help agents listen to sound, judge it, and make it themselves. It is written in Python and provides a CLI, a Python API, and an MCP server. MCP is the standard interface coding agents use to connect external tools. It runs on CPU alone and is published under the MIT license. The latest release is v0.2.0, and the materials contain no changelog for that release, so what was added cannot be pinned down.
Listing the features shows two directions: listening and making. The center is describe, which runs the public CLAP embeddings and AudioSet tagger on CPU to describe sound in words; match, which scores how well a description fits the sound; and find, which locates a suitable segment in a long recording. It separates real recordings from sound synthesized in code, and it produces a time-by-time listening log for a whole video, annotated images, and a score table. There is also a function to evaluate the sound of launch videos. earshot forge makes sound in code, and earshot fit turns a recording into editable code. The eval directory holds blind evaluation and benchmark scripts, the tests path has 29 entries, and documentation covering latency and limits is published as well. Stars stand at 2027.
However, most of the scores shown are results of its own benchmark, so they cannot be read as external validation. The technical score is also set low, at 7 out of 10. It is usable as a tool for giving agents hearing, but there is still time to wait before its scores can be weighed against external criteria.