Jev Skill Gate

Partly, and the README admits it. The free local TF-IDF scorer reached mean pairwise AUC of 0.961 over 18 cases and still hid 5 labeled skills, all at exactly 0.00, on prompts sharing no vocabulary with those skills. The author's own read is that if your descriptions share words with how you phrase prompts, the free scorer is close enough and you can skip the key.

Thinner than the table looks. Only 4 of the 20 labeled cases ran against Jev, because the free tier allows roughly 4 requests per refill window and nothing is estimated to fill the gap, against 217 installed skills. Surviving is also not full visibility: 5 of 19 expected skills landed name-only, giving Claude the name and no description to judge it by.

It is an estimate, not a measurement. The README notes that /context reports 9.9k for skills on that machine against the 12,750 the estimator counts, because it bills characters at 3.8 per token and runs slightly high; the claim is that the savings ratio transfers, not the absolute number. Computing it costs about $0.00074 per session and is cached for seven days.

No, by design. It maps scores onto on, name-only and user-invocable-only and never writes off, so a wrong call costs tokens you wanted to spend rather than access, since hiding from Claude is reversible by typing the slash command and hiding from you is not. Installing does, however, register a SessionStart hook in your Claude settings, merged into whatever hooks you already had.

Requires Node 18+ with no dependencies. Clone the repo and run `node bin/jev-skill-gate.mjs install`, which registers a SessionStart hook in ~/.claude/settings.json and merges into any hooks you already have. The config file it writes is mode 0600.

Yes. Without a key it uses a built-in local scorer and still works; you can optionally set AI_GATEWAY_API_KEY for Vercel AI Gateway or TYPESAFE_API_KEY for TypeSafe direct, or store one in the config file. An environment variable always wins over the config file.

It maps a relevance score onto on, name-only or user-invocable-only. It never writes off: hiding a skill from Claude is reversible by typing its name, but hiding it from you as well is not, so a wrong call costs tokens you wanted to spend rather than access.