Stuffing 40 skills into the system prompt taxes every token. Skills inverts this — agents discover the right one at runtime, fetch the folder, run it, rate it. The catalog gets sharper without you curating.
Stop curating skill libraries. Agents find the right one at runtime, fetch it, run it, and rate it.
Your skill list is whatever your platform has indexed. The agent calls find_skill the moment it needs a capability. Hybrid search returns the top matches with review aggregates. The agent picks. You don't.
After every run, the agent leaves a 4-dimension review. Accuracy, value, efficiency, clarity. The composite boosts what works on the next search. New skills get a fair shot — the boost is multiplicative, never punitive.
Stuffing 40 skill descriptions into a system prompt costs ~12k tokens per call and distracts the model. Skills (the product) ships one MCP tool — find_skill — and lets the agent pull only what it needs.
Stuffing 40 skills into the system prompt taxes every token, distracts the model, and rots into a museum nobody uses. Runtime discovery is the fix — but it needs a registry, content-hash dedup, agent reviews that compound, and a fetch path fast enough to call inline. We did all of it.
We watched five AI startups in a row hand-roll the same skills stack — and burn three weeks doing it. We packaged ours so you don't have to. Drop the SDK in once; this product, plus the rest of the suite, comes with it.
Admin-vetted GitHub repos sync every 30 seconds. Plugin-bundle dedup collapses byte-identical copies. Your agents see one match per unique skill, not seven.
Upload via API. Owned by the platform key (sk-plat_), encrypted at rest in GCS. Visible only to your agents, never the public catalog.
Created with sk-eu_, owned by the end-user — not the key. Survives API rotation. Each end-user gets their own evolving capability set.
Best-rated skill wins on the next search. New skills get a fair shot via multiplicative boost — never punished for having no reviews yet. Failures stay visible, not invisible.
find_skill({ query: "summarize PDF" })fetch_skill(skill_id, target_path)
// runtime materializes + executes
submit_skill_review({ accuracy, ... })We stopped maintaining a curated skills list. The agent finds what it needs, runs it, and our system-prompt size dropped 70%.
— Adam · CTO, Quillper request, after free tier. No markup on tokens. Cancel anytime.
We onboard 1–2 indie startups a week. If you'd rather ship features than maintain a skills stack, talk to us.