earshot
earshot is an open-source Python CLI and MCP server that lets coding agents describe, match and critique audio, and generate sounds from code.
No visits from OpenBeat in the last 30 days yet. Every click on “Visit” is counted here.
Pricing and API checked by an OpenBeat editor on Nov 11, 2026 · facts from the maker’s own site ·
Tool insights
earshot lets a coding agent analyze audio files and judge whether a sound fits a text description. It also lets the agent build sounds in code and check the result by listening models, all locally on CPU.
Coding agents work with text and code, so they cannot hear audio they are asked to analyze or produce. earshot gives them measurable tags, verdicts and reports so they can work with sound.
Developers building with coding agents such as Claude Code who need audio analysis or code-generated sound effects, for example for launch videos. People who want a hosted, instant-response service or a polished GUI should look elsewhere.
What it does
earshot is an open-source tool by Leonxlnx that gives coding agents a way to perceive sound. An agent can ask it to describe a recording, check whether a sound matches a written description, or build a new sound from code and judge the result. It runs as a command-line tool, a Python package and an MCP server, and it works on CPU.
The listed models are CLAP and AudioSet labels. Gemini is optional, through the llm command.
Key features
- Describes sound with calibrated tags and measurements.
- Matches a sound to a text description and returns a verdict.
- Finds the best moments inside long recordings.
- Checks whether audio is real or synthesized in code.
- Renders listening transcripts and annotated spectrogram sheets.
- Scores launch-video sound against professional reference films.
- Makes sounds in code from recipes, judged by listening models.
- Translates a recording into an editable synthesis recipe.
- Ships a bank of 252 sound effects that can be rendered by id.
Pricing
earshot is open source under the MIT license. As of November 2026, the facts list no paid plans and no price. Whether a separate free plan exists is not stated, since the tool is simply free to download and run. The optional llm command needs your own Gemini key, so any cost there depends on your Gemini account.
API and integrations
The tool has a Python SDK and a CLI. It also runs as an MCP server and lists Claude Code as an integration. The code is on GitHub at Leonxlnx/earshot. A separate docs site is not stated on the maker's site, but a changelog is kept in the repository.
Good to know
- Each CLI call loads models first, which takes about 20-40 seconds.
- Reference recordings for the forge --refset option are not included in the repo.
- The optional llm command needs your own Gemini key and your explicit consent.
- The author notes that the realness model is easy to fool, so treat its verdicts as hints, not proof.
- Recent releases: 0.4.1 (6 November 2026) fixed NaN/Inf sample errors, set BLAS thread limits and added a uv install. 0.4.0 added 51 new sounds, a launch kit preset and a rebuilt fit engine. 0.3.0 added the sound bank, sfx renders and film kits.
- Founding year and company details are not stated on the maker's site.
FAQ
Is earshot free?
Yes. earshot is open source under the MIT license, and as of November 2026 no paid plans are listed. The optional llm command needs your own Gemini key, which is billed by Google, not by earshot.
Does earshot work with Claude Code?
Yes. earshot lists Claude Code as an integration and runs as an MCP server. It can also be used from the command line or as a Python package.
Does earshot need a GPU?
No. earshot runs locally on CPU. Each CLI call loads its models first, which takes about 20-40 seconds.
Can earshot tell if audio is real or AI-generated?
It has a check for whether audio is real or synthesized in code. The author notes that this realness model is easy to fool, so results should not be treated as proof.
What can earshot generate?
It makes sounds in code from recipes, judged by listening models, and can turn a recording into an editable synthesis recipe. It also ships a bank of 252 sound effects that can be rendered by id.
Releases
No tracked releases yet.
In the news
No coverage yet.