[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2ret79xcflvxq":3},{"code":4,"message":5,"data":6,"timestamp":31},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":16,"summary_md":17,"topic":18,"github_topics":21,"license":27,"homepage":28,"updated_at":29,"history":30},"Leonxlnx\u002Fearshot",8,"Ears for coding agents: let AI hear, review and make sound. CPU-only, open source.","Python",2630,160,270,9,[],"Lets coding agents analyze, judge and synthesize sound and launch-film audio locally on CPU.","## The problem\n\nCoding agents can generate video, product demos and games, but they have no way to evaluate the audio that goes with them. Choosing sound effects, checking whether music swells land on the right cut, or judging whether a sound feels \"premium\" all require human ears. earshot gives an agent measurable audio feedback: text descriptions, fit scores, moment-finding within long recordings, sync checks against picture cuts, and an annotated image of the audio timeline.\n\n## What's different\n\nThe usual route is either a hosted audio-LLM API (accurate on semantics, expensive, slow, requires uploading media) or raw DSP libraries that return numbers an agent cannot interpret without a human. earshot sits in between: it runs a fixed local pipeline on CPU (no GPU, no network by default) and phrases results in calibrated language an agent can act on, including z-scores against a vocabulary and concrete fixes with timestamps. Its second half is unusual: `forge` synthesizes new sounds from code recipes and improves them through a critic loop, and `fit` turns a recording into a re-renderable recipe. The README is also candid about latency — each CLI invocation reloads models in 20–40 seconds, so the MCP server or the Python API is the intended path for repeated calls.\n\n## Who it's for\n\nDevelopers building agent workflows around motion graphics, game audio or product films, where an agent needs to iterate on sound without a human listening to every render. Teams that must keep media on-premises will value the CPU-only, local-first design. People who need broadcast-grade audio mastering, real-time playback processing, or a DAW replacement should look elsewhere — this is an evidence and synthesis tool, not an audio editor. Anyone without an agent in the loop will find the CLI verbose.\n\n## Cost of adoption\n\nPython package, MIT licensed, CPU-only. There are model and data files to install, and `doctor --load` exists to verify them. Two real costs: the 20–40 s model load per CLI call (except `measure`, `sfx`, `kit`), and the conceptual overhead of learning the critic's vocabulary and the recipe format. The optional Gemini upload path requires your own key and explicit consent. The project is 9 days old at 0.4, so APIs and recipe formats may still move.\n\n## Verdict\n\nThis is a brand-new project riding a fast spike: 2,627 stars in about nine days and roughly +274 per day, with no Hacker News discussion and no daily trending board appearance. Those signals suggest launch-week attention rather than settled adoption. The problem it targets is genuinely underserved — agents that make media cannot currently hear it — and the local, CPU-only, MIT-licensed approach is the right shape for that gap. But at 0.4 with a README that already documents a v2 rework of its core benchmark, expect churn. Worth a trial if you are building audio-aware agent pipelines today; wait a few releases if you need stable interfaces.",{"slug":19,"name":20},"media","Media & Design",[22,23,24,25,26],"ai-agents","audio","claude-code","mcp","sound-design","MIT",null,"2026-10-24T23:35:49Z",[],1792901833]