Price terms as published

Each line is the vendor’s own wording, split from the price capture checked on 2026-10-24. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.

2 price terms for earshot
#Plan / termChecked
1Free (MIT open source)2026-10-24
2optional Gemini-based critique needs your own API key2026-10-24

One-liner: earshot is “ears for coding agents” — a local, CPU-only toolkit that turns sound into something an LLM can reason about. Agents that write code for films, product demos, or games still cannot hear what they made; earshot gives them calibrated descriptions, searches, timelines, and scorecards, plus the ability to make their own sounds in code (earshot forge) and translate a recording into editable code (earshot fit).

Everything below was checked against the official repository (GitHub API + raw README, fetched 2026-10-24) and two independent trackers/coverage items.

Verified facts (GitHub API + raw README + third-party, 2026-10-24)

Attribute Value Source
GitHub stars 2,489 (155 forks) Leonxlnx/earshot via GitHub API, 2026-10-24
License MIT (code); models downloaded under their authors’ own licences GitHub API + README license note
Created / last push 2026-10-15 / 2026-10-22 GitHub API
Language / stack Python 3.10+, ffmpeg/ffprobe on PATH; CPU only, no GPU README “Install”
Interfaces CLI, Python API (from earshot import api), MCP server for Claude Code and other agents README “What it does” + “MCP server”
Core commands describe / match / rank / find / compare / ab / realness / timeline / sheet / critique / profile / measure README command table
Sound creation forge RECIPE (make a sound in code, improve by listening), fit FILE (translate a recording into a forge recipe), sfx render / kit (252-code sound bank with sync-point measurements) README “What it does” + “New in 0.4”
Critique critique film.mp4 scores a launch video against a profile of 7 professional launch films, with timestamped “fix first” items README command table
Models CLAP (laion/larger_clap_general @ pinned revision) + AudioSet tagger (MIT/ast-finetuned-audioset-10-10-0.4593), ~1.1 GB total, downloaded once into repo .cache/hf (or EARSHOT_HOME); offline afterwards README “Install”
Offline EARSHOT_OFFLINE=1 / HF_HUB_OFFLINE=1 / TRANSFORMERS_OFFLINE=1 — never downloads, missing model = named error README “Install”
Startup cost Each CLI call loads models in ~20–40 s (measure/sfx/kit load none); MCP server or one Python script keeps models loaded README “What it does”
Optional paid piece llm critique (Gemini) reads GEMINI_API_KEY from env, explicit --yes-upload consent; never write a literal key into a file README command table + MCP section
Install git clone + venv + pip install torch --index-url .../cpu + pip install -e ".[mcp,llm,dev]" + earshot doctor --load README “Install”

This is an official-terms entry: we have not installed earshot or run its models locally, so there is no score and no hands-on claims.

Getting it running

Installation is a Python project: clone the repo, create a venv, install CPU-only PyTorch from the official CPU index, then pip install -e ".[mcp,llm,dev]" and verify with python -m earshot doctor --load. The first model use downloads two open models (~1.1 GB total) into the repo’s .cache/hf (or EARSHOT_HOME): CLAP for text-audio embeddings and an AudioSet tagger for sound-event labels. After that it loads from cache only and never contacts the Hub; earshot doctor --fetch pre-downloads them, and EARSHOT_OFFLINE=1 makes any missing model a named error instead of a silent download.

For agents, the recommended path is the MCP server: claude mcp add earshot --scope user -e EARSHOT_THREADS=3 -- /path/to/earshot/.venv/bin/python -m earshot mcp, or an .mcp.json entry for any client. Models load on first call and stay warm. The optional llm_critique tool only activates with your own GEMINI_API_KEY and explicit --yes-upload consent — the README is explicit that you never write a literal key into a file. Two agent skills also ship in the repo (skills/earshot for how to listen, skills/real-sound for premium soundtracks from licensed recordings).

What it costs

The software is free, MIT, and runs fully locally on CPU — the README states no GPU is used, and every command except the optional Gemini critique works offline with no network and no paid API. The two costs to plan for: one-time model download (~1.1 GB) and the compute time of model loading per CLI call (~20–40 s, or keep models warm via MCP/Python API). The only genuinely paid piece is optional: the audio-LLM critique (earshot llm) requires your own Gemini API key and is billed by Google. Reference collections for fit and forge --refset are licensed recordings that are not shipped in the repo — bring your own (the README says the repo contains no audio or video files).

How it differs from alternatives

  • Jockey by TwelveLabs (this site’s entry) is a hosted video-intelligence API: cloud, paid per video, focused on understanding footage. earshot is the opposite pole: local, CPU-only, free, and aimed at sound in code-made films rather than scene understanding.
  • Whisper-class transcription tools transcribe speech to text. earshot’s describe/match/find are calibrated audio perception — what a key press sounds like, whether a description fits a sound, where inside a long recording a sound occurs — plus realness (real recording vs code-synthesized) and timeline/critique (sync sound to picture, score against professional launch films). It does not transcribe speech as its primary job.
  • Cloud SFX libraries / stock sound provide finished files. earshot’s forge/sfx/kit generate sounds in code deterministically — no models, no network, rendered by id in seconds, audited against the picture with measured onset/peak/end — and fit turns a recording into a variable recipe, so the same sound can be re-rendered with knobs.

Where to stay honest

  • Per-command latency is real: ~20–40 s model load per CLI invocation; the README itself recommends the MCP server or a single Python script for more than a few calls.
  • ~1.1 GB first download is required before any model-based command works (measure/sfx/kit work without models).
  • No reference recordings ship in the repo — fit/forge reference sets need your own licensed clips; the README keeps the evaluation data (licensed recordings, third-party films) out of the repository and asks you to keep the gallery page private for the same reason.
  • The Gemini critique is a third-party add-on, not core earshot, and uploads audio only with explicit consent and your own key.

Who it’s for

  • Developers running Claude Code/Codex who make videos, demos, games, or UI sound effects in code and want the agent to hear, verify, and improve the result
  • MCP-first agent builders looking for a local capability with no cloud dependency
  • Sound-design-curious tinkerers who want an agent to forge whooshes, hits, and key clicks deterministically and then critique its own film
  • Not for: speech-transcription workloads (use Whisper-class tools), teams that cannot accept a 1.1 GB local model cache, or anyone who wants audio understanding without installing a Python toolchain

Sources