earshot gives coding agents ears: hear the soundtrack, score the sync, forge SFX in code
Agents can now render a whole launch film but can't hear it. earshot is a CPU-only CLI, Python API and MCP server that turns sound into evidence: what a file sounds like, where the hits land against the cuts, and a scorecard with timestamped fixes.
Last week Claude Eyes gave agents a way to watch their renders. earshot covers the other half. Its README puts the problem plainly: an agent can't tell a premium key press from a cheap one, notice that a hit lands 80 ms after the cut, or hear that the music swells in the wrong place. Leon Lin (@LexnLin) created the repo on Oct 15 under the MIT license. His launch post passed 120 likes, a star-velocity tracker flagged it at about 290 stars a day this morning, and @DanKornas picked it up today. It had 1,494 stars when we checked.
The parts video makers will use: describe says in words what a file sounds like. match scores how well a sound fits a description against named alternatives. find locates the best moments inside a long recording. timeline produces a second-by-second listening transcript with events, loudness, a beat grid, and how the sound syncs to the picture cuts. sheet renders all of that as an annotated PNG, video frames included, for agents that read images. critique scores a finished video against a profile built from 7 professional launch films and returns timestamped fixes. You can build your own profile from films you admire with profile.
It also makes sound. forge has the agent write a sound as a physical-synthesis recipe, renders variants, scores them against real reference recordings, and keeps improving them by listening. Version 0.3 adds a bank of 201 code-made effects (whooshes, hits, clicks, risers, logo stings and more). They render by id in seconds with no models and no network, and sfx and kit aim them at exact moments in the picture and report the measured sync points. The README says every sound in the 22-second launch film was made this way, with no recordings and no music track.
Practical notes from the README: each CLI call loads models for 20 to 40 seconds, so for more than a few checks run it as an MCP server or from one Python script. measure, sfx and kit need no models. There's an optional Gemini critique, but it only runs with your own key and explicit consent to upload. Our take: pair earshot's timeline and critique with a Claude Eyes or contact-sheet loop, and an agent can check both picture and sound before you watch. Spotted, not tested. Credit Leon Lin.
Skills & repos in our directory
- earshot · MCP, Audio & music
- claude-motion-explainer · Skill, Motion graphics
- Retake · MCP, Code-to-video engine
- whiteboard-teacher-video · Skill, Motion graphics
- claude-codes-influencers · Repo, Lip-sync & timing
- VoiceStudio · MCP, Audio & music
More from the Lab
- huashu-motion: an open skill that writes the launch film, scores it, and lets you point at the frame to fix itSpotted, not yet tested
- claude-motion teaches an agent taste, then makes it check its own framesSpotted, not yet tested
- BaoCut: an AI video agent where every caption, dub and cut stays editableSpotted, not yet tested
- Tandem: an MIT video editor with a CLI and MCP, where agent edits stay purple until you've watched themBased on the creator’s video · not yet tested by us
Topics: Claude Motion & code motion skills · Audio, voice & music · Video MCP servers