MCP-native skill registry · ARD-adjacent
Where agents find skills — and you find what they used.
Author multi-file skills, serve them over one MCP endpoint your agents read at runtime, and measure exactly what got discovered, loaded, and ignored.
works withClaude CodeCursorCopilotLangChain
contextpie — skillsMCP active
798
Loads · 30d
1
Dead skill
12
Discovery gaps
| Skill | Last | Loads |
|---|---|---|
| react-component | 3h | 412 |
| pdf-extraction | 12m | 248 |
| git-commit-style | 1h | 96 |
| legacy-deploydead | — | 0 |
find_skills(...) ↳ top score 0.86 ↳ get_skill ↳ get_skill_file ✓ logged · search → load → file_load
Build → Use → Measure → repeat
01BUILD
Authoring studio
Write multi-file skills with AI-assisted, discoverability-first authoring.
02USE
Hosted MCP server
Agents connect once and discover, load, and fetch skills at runtime.
03MEASURE
Analytics that close the loop
Every call is logged — see what's used, what's dead, and where discovery failed.
The query is the telemetry.
When an agent queries our MCP server to discover and load a skill, usage data lands in our logs as a byproduct of serving the request. No proxies, no wrappers.
searchloadfile_load
Build
Discoverability isn't metadata. It's the whole game.
- Description + trigger phrases are what get embedded and searched — they decide whether a skill is ever found.
- Inline quality nudges keep them distinct and well-scoped.
- Draft with AI gives you a strong first pass to refine, not a black box.
✦ Discoverability
Description ● distinct
Extract clean text and well-formed tables from PDF documents with an OCR fallback.
extract tables from a pdfocr a scanned doc
✦ Discovery gaps
mock api responses for testsno match
write terraform for gcp0.38
convert figma frames to react0.41
Measure
Three truths nobody else can show you.
- What's used — and what's dead weight you can deprecate.
- Where agents searched and found nothing good — your to-do list.
- Which bundled files never get fetched.
Progressive disclosure, by design
find_skills()Returns lightweight descriptors — name, description, score. No bodies, no files.
get_skill()Loads the SKILL.md body plus a manifest of supporting files.
get_skill_file()Fetches a referenced file only when the SKILL.md asks for it.