earshot: a local MCP server that gives coding agents ears
Created October 15, MIT-licensed, 679 stars when I checked on October 18: earshot describes audio in words, scores how well a sound fits a description, finds a moment inside a long recording, critiques a film's sound against professional references, and synthesizes effects in code. It runs locally on CPU as a CLI, a Python API and an MCP server. I read the repo and the LICENSE; I did not install it.
Facts checked October 18, 2026
earshot is a new open-source tool by Leon Lin, created on GitHub on October 15, 2026, that gives coding agents a way to reason about sound. It describes what an audio file sounds like in words, scores how well a sound matches a description, locates a moment inside a long recording, critiques a film's audio against seven professional launch films, and can synthesize effects in code with earshot forge. It runs locally on CPU and ships as a CLI, a Python API and an MCP server. The license is MIT, which I read in the repo. I installed and ran nothing. Repository
What it does
The README frames the gap plainly: agents can now write the code that renders a film, a product demo or a game, but they cannot hear what they made. earshot turns sound into evidence an agent can reason about. The documented commands: earshot describe produces a word picture with measured attack, decay, peak and brightness; earshot match scores a file against a description with z-scores against a 322-description vocabulary; earshot find locates matching moments inside a long recording; earshot critique scores a film against seven professional launch references and returns a fix-first list.
Two generative directions complete the loop. earshot forge synthesizes sound effects in code, and earshot fit translates a recording into code the agent can vary. A secondary write-up reports fit arrived in v0.2 on October 16; I could confirm the commands exist in the README but not the version date, so treat that date as reported, not verified.
The MCP setup, as documented
The README gives a one-line install for Claude Code: claude mcp add earshot --scope user -e EARSHOT_THREADS=3 with the project's virtualenv Python running -m earshot mcp, or the equivalent block in any client's .mcp.json. Models load on the first call and stay warm, taking about 20 to 40 seconds, and the README advises starting Claude Code with MCP_TIMEOUT=60000 on a slow machine.
The stated hardware story is modest: everything runs locally on CPU. The repo is three days old and single-author. It showed 679 stars and 39 forks when I checked on October 18. New repos can gain stars quickly for reasons other than use, so treat the count as a name to watch, not an endorsement.
Privacy and limits
The core promise relevant to privacy is local execution: the analysis tools run on your CPU, so audio does not need to leave the machine for describe, match, find or critique. The install extras include an optional llm group; I did not trace what, if anything, that path sends to a network, so the local claim is verified for the core tools only.
Maintenance is the other honest limit: one author, created October 15, zero open issues because the project is days old. The MIT license permits anything, including you fixing it yourself, which matters more than usual for a tool this young.
Why this category matters
The pattern to notice is the closing of a sense loop. Screenshot tools let agents see their UI output; earshot is the same move for audio: generate, listen, score, fix, all inside the agent's reasoning. Film demos, games and product videos are exactly where coding agents have been producing silent, unreviewed sound.
If the approach holds, expect a family of sense-tools for agents: small, local, CPU-friendly evaluators that turn non-text output into text an agent can act on. The MCP server shape means any client that speaks the protocol gets the capability for free.
The documented MCP and CLI surface
# .mcp.json (any MCP client), adapted from the README:
{
"mcpServers": {
"earshot": {
"command": "/path/to/earshot/.venv/bin/python",
"args": ["-m", "earshot", "mcp"],
"env": { "EARSHOT_THREADS": "3" }
}
}
}
# CLI examples quoted from the README:
earshot describe enter.ogg
earshot match key.ogg "a deep thocky mechanical keyboard press"
earshot find keys.ogg "a single mechanical key press" --top 4
earshot critique film.mp4
# On a slow machine: MCP_TIMEOUT=60000 - models load on first call (20-40 s).