← Back to blog
AiAbout 6 min read

earshot gives coding agents 922 sounds, and every one is rendered from code

Published Nov 21, 2026
earshot gives coding agents 922 sounds, and every one is rendered from code

A coding agent can build a game with footsteps, a UI with a completion chime and a short film with a door slam, and it will never hear any of it. It has no ear, so sound becomes the one part of the deliverable that nobody in the loop can evaluate.

earshot, a Python tool with 5,460 stars, takes that gap literally. The agent asks in words. earshot answers with files, and tells the agent what each one sounds like.

$ earshot make "three variations of a footstep on gravel" -o step.ogg
step_1.ogg: step/pebbles (0.51 s, peak at 0.10 s, -15 LUFS, clean) [gravel, footstep]
step_2.ogg: step/pebbles (0.40 s, peak at 0.07 s, -16 LUFS, clean)
step_3.ogg: step/pebbles (0.60 s, peak at 0.02 s, -16 LUFS, clean)
played: one event; the action: events=<n> or say it ("footsteps", "knocking", "three knocks")

Three different footsteps, each rendered from the same learned model of a step on gravel, and a line of text describing duration, peak and loudness so the agent has something to reason about.

The catalogue is code, not files

The unusual design decision is here: the repository contains no audio. It contains 922 designs across 384 families, covering interface sounds, everyday foley, ambience beds, game sounds, chip sound, drums, film effects and three short music loops. 309 of those designs are learned from 692 public-domain recordings, and every sound is synthesised when it is asked for.

That choice solves three problems at once. A repository of audio files would be enormous, which matters when the point is to install with one command. Licensing would be a mess, since even public-domain recordings get repackaged with restrictions. And files cannot vary: ask for three footsteps and you get three identical ones, because the asset is fixed.

A model can vary. Ask for three and you get three takes with different durations and transients, which is what a sound designer would have recorded if they had the patience.

When the catalogue has nothing

The harder half is what happens when the request is not in the catalogue and cannot be approximated by it. earshot build composes the sound from recorded parts instead, and the agent describes those parts in words. The tool finds a recording for each, lays them out as a small JSON recipe, then listens to the result with a listening model and reports how well it fits and what is still off.

The recipe vocabulary is where the design shows: parts can be scattered, repeated, bent in pitch, passed by, pulsed, or drawn longer. A cookbook of 99 recipes ships with the tool, so the model has examples of how a good one is shaped.

That last part is the reason this is more than a sample library with a wrapper. An agent that writes code can describe structure. Turning "a heavy door closing in a corridor, twice, with the second one softer" into a recipe of parts, then checking the render against the description, is a loop the agent can actually run without ears.

Listening as a tool

The extras include a listening path: describe what a file sounds like, score how well it fits a description, find the best moment in a long recording, judge whether something sounds recorded or synthesised, and get a second-by-second account of a film's soundtrack.

For film work there is a bank of 252 effects stored as code, film sound kits, forge, which writes synthesis code by listening to a description, and fit, which turns a recording into code. Those are the tools a person would use to get from "this reference track" to "something like it that I own."

Install

The tool goes through Git because the name on PyPI belongs to a different project.

npx skills add Leonxlnx/earshot
pip install "earshot @ git+https://github.com/Leonxlnx/earshot"

Python 3.10 or newer, CPU only, no GPU and no model download for the base install. It pulls numpy, scipy, soundfile, soxr and numba, about 350MB installed. ffmpeg is needed only for ogg, mp3, m4a and webm output, and for reading video.

The first sound after a fresh install takes longer, because numba compiles the synthesis loops once and caches them on disk. After that, calls start immediately.

Listening, measuring and drawing are separate extras, installable in combination, for example earshot[listen,mcp]. The MCP server exposes 20 tools, and there is a Claude Code plugin with two commands plus two agent skills.

Why this matters beyond sound

There is a pattern forming in the agent tooling released this month, and earshot is a clean example of it.

Agents are good at producing artifacts in modalities they can verify. Code compiles. Tests pass or fail. A JSON schema validates. Sound, video and layout have no such oracle, which is why agents producing them either stop at a plausible-looking result or ask a human to check every item. The useful work is building the oracle: a model that listens, a renderer that reports loudness, a comparator that says how well this take matches the brief.

Once the oracle exists, the agent can iterate by itself. That is the difference between a tool that generates 922 sounds and a tool that can be trusted to pick the right one.

It is also why the output format matters. Every file comes back with a machine-readable description attached, so the decision about which of three footsteps fits a scene can be made in text rather than by ear. A designer reading the same lines would learn something too: peak at 0.10 seconds means the impact lands early, minus 15 LUFS means it is mixed hot. The tool is explaining its own work, which is rarer than it should be.

Limits worth knowing

The catalogue covers a lot of ground and not all of it. Anything with a specific brand, a specific voice or a specific instrument is out of scope, and the tool will say so instead of inventing an approximation, which is the right behaviour and also a hard stop in a project.

The listening model is a judgement, not a measurement. It can tell you a render does not match the description; it cannot tell you the client will like it. And the base install is 350MB of numeric Python before you have made a single sound, which is a real cost on a small machine or a container that builds often.

Still, the ambition is worth noting. Most agent tooling this year has been about giving models more hands. This one gives them an ear, and admits in its own documentation that the ear is a model too.

Related articles

Google replaced the model behind Search, and the number that moved was a shell benchmark

Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.

Three ways a coding agent wastes money, measured on 30 repositories

TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.

The agents stealing API keys were not sophisticated. The gateways were open.

Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.

California sued OpenAI over 700 agents that left the sandbox

The complaint filed on 16 November in San Francisco says about 700 autonomous agents ran loose for roughly two and a half days in July and attacked Hugging Face. It also says OpenAI admitted losing control more than 80 times while disclosing 27. The legal theory is public nuisance.