Earshot gives AI coding agents ears to tweak their own sound effects

Earshot is an open-source tool that basically gives AI coding agents ears. It lets agents like Claude Code not only synthesize sound effects in code, but actually listen back, critique the result, and rewrite the code to fix it.

AI agents can already write the code to render product demos, UI animations, and game clips. The glaring bottleneck is that they are deaf. An agent cannot tell a premium mechanical keyboard thock from a cheap plastic click, notice a punch sound hitting 80 milliseconds late, or hear background music swelling over a cut.

Earshot turns audio into structured evidence an LLM can reason about. It translates audio files into plain-language descriptions, checks acoustic match against a target prompt, pinpoints specific sound events inside long recordings, and generates a second-by-second listening transcript with annotated spectrograms.

How it actually works

Instead of relying on heavy black-box audio diffusion models, Earshot treats sound as procedural code.

The project ships with a built-in bank of 252 sound effects synthesized purely through physical code algorithms. These cover rounded UI pops, wood knocks, camera shutters, reverse cymbals, short risers, and glitches. Because they render mathematically on the fly, generation takes seconds without internet access or GPU infrastructure.

The real power lies in the iteration loop. An agent drafts a synthesis recipe using contact excitations and modal resonators. Earshot renders candidate variants, and an internal critic scores them on tonal balance, temporal envelopes, and artifacts like clipping or harsh ring. The agent reads the critic's plain-text feedback, tweaks the parameters, and re-renders until the sound lands right.

Hardware and getting started

Earshot runs entirely locally on CPU, so there is no need for high-end GPUs or cloud API keys.

You can run it as a standalone CLI, import it into Python scripts, or hook it directly into coding setups via its Model Context Protocol server. Standalone CLI calls take roughly 20 to 40 seconds to cold-boot the evaluation models, so keeping the MCP server running is the recommended path for interactive agent workflows.

The repository is publicly available on GitHub under the Leonxlnx namespace. Recent updates also introduced launch-kit presets that automatically balance volume levels against background music beds and snap whooshes cleanly around video cuts.

Why it matters

If you use coding agents to build UI animations, product teasers, or small games, Earshot tackles one of the most frustrating parts of the pipeline: endless manual trial and error just to line up a whoosh.

We like this approach because it skips the compute overhead of generative audio models in favor of lightweight physical synthesis and programmatic feedback. It is strictly built for short, crisp functional sound effects, so do not expect it to score a cinematic narrative, but for agent-driven media workflows, it is a very clever step forward.