Earshot lets AI coding agents listen to a launch film

An AI coding agent can write the code that renders a product film, a demo or a game. Until this week it could not tell whether the whoosh at the cut sounded cheap, the key press landed on the wrong beat, or the music swelled in the wrong place. A new open-source project called earshot, from the GitHub user Leonxlnx, gives the agent ears: a CLI, a Python API and an MCP server that score a sound file against a brief, find the best moment inside a long recording, and critique a finished film against seven professional launch videos.

The repository was created on 15 October and has 654 stars and 39 forks on GitHub two days later, under the MIT licence. It is CPU only, with no GPU required; the models are around 1.4 GB on first download and cached locally for later calls.

The eleven commands

The project lists eleven commands. describe returns what a file sounds like in calibrated tags, with the nearest alternative readings and measured peak, brightness and envelope. match scores a file against a text brief and returns a verdict of strong, plausible, weak or no match. find returns the best moments inside a long recording as sample trims. realness labels a sound as a real recording or code-synthesized, with per-window evidence. critique scores a whole film against seven professional launch films and prints timestamped fixes, with lines such as "nothing reaches body level until 0:01.01". There is also a sheet command that renders an annotated PNG of frames, labelled events, spectrogram, waveform, loudness and sync markers, so an agent that reads images can look at the film instead.

Two further commands connect listening and writing code. forge renders a sound in code from a recipe, scores it with earshot's own critic, mutates the recipe and runs again. fit takes a recording and writes a Python recipe module that re-renders it, which the agent can then vary. The same tools are reachable from Claude Code and other agents through an MCP server, which keeps the models loaded between calls.

What the measurements prove, and what they do not

The repository is clear about what the measurements mean. In a blind test the author picked the code-synthesised sound correctly in eight of fifteen trials, which is about what random guessing would give, with one listener and fifteen tries. The description model's "strong" verdicts match the sound type 98 percent of the time, and the real-against-synthetic classifier reaches ROC-AUC 0.92; both numbers are the author's own on earshot's own held-out set. The forge critic uses earshot's own models, so a search can please them without the sound getting better, and the README names that risk as a Goodhart problem. The reference recordings used by the critic are licensed and not inside the repository; a team has to bring its own.

Processing a 60 to 90 second film takes about 1.5 minutes on a laptop CPU. Each CLI call is a new process and loads the models in 20 to 40 seconds, so a workflow that makes a dozen calls should use the MCP server or a single Python script.

A team releasing a video an AI coding agent produced now has one more step before the file leaves the laptop: ask the agent to listen to its own work. The answer the agent gets back is measurements and words, with every suggested fix tied to a timestamp.

Source

Leonxlnx/earshot, MIT licensed. The launch film and its listening sheet are in the v0.1.0 release.