Retrieve skills
at runtime.
Not upfront.
https://skill-mcp.jigneshponamwar.workers.dev/mcpStatic skill files
don't scale.
Every approach before skills-mcp forces the same trade-off between agent capability and context efficiency. The model gets everything or nothing.
context window exhausted before first user message
Context window exhausted upfront
Embedding 30 expert procedures into every system prompt costs 60–80k tokens before the agent sends a single message. Most of that context is never used.
Copy-paste maintenance across projects
Every project gets its own skill file copies. One update means manually syncing N files across M projects, with no version control and no diffing.
Reconfiguration per AI platform
Claude Code, Cursor, Windsurf, Cline: different config formats, different file paths, different scopes. The same procedure files multiply into incompatible formats.
No relevance matching
Static files can't score themselves. No way to retrieve the procedure best suited to the task at hand. It's all procedures or none.
What if the agent could request exactly the skill it needs, the moment it needs it?
Three tiers.
Loaded only when
needed.
skills-mcp serves knowledge in layers. The agent always starts with the lightest context and escalates only when the task demands it.
Semantic skill search
The agent calls skills_find_relevant() with a plain description. Qdrant returns the highest-scoring match by cosine similarity against embedded descriptions.
skills_find_relevant("stripe webhooks")
→ 0.97 stripe-integration [tier-1]Returns frontmatter only. ~200 tokens.
Full procedure on demand
When ready to act, the agent calls skills_get_body(). The complete step-by-step procedure is returned for exactly one skill, not all 30.
skills_get_body("stripe-integration") ← latest
skills_get_body("stripe-integration", version="1.2") ← pinned
Returns a deprecation_notice if the skill has been superseded.Full procedure + patterns. ~2–4k tokens.
Deep references on request
For complex tasks, tier-3 tools pull curated docs, API references, and scripts. Only loaded when the skill body explicitly references them in the manifest.
skills_get_reference("stripe-integration", "WEBHOOKS.md")
→ Verification spec · Security checklistOnly loaded when the body references it.
Try it in 60 seconds.
My hosted instance is live. Connect it to your AI agent right now.
{ "mcpServers": { "skill-mcp": { "transport": "streamable-http", "url": "https://skill-mcp.jigneshponamwar.workers.dev/mcp" } } }
Paste this config into your AI tool’s MCP settings. The server is already running no deployment needed. Per-IP rate limiting: 60 requests/min (configurable via RATE_LIMIT_RPM). Exceeding the limit returns HTTP 429.
30 production-ready
expert procedures.
Across 7 categories: API integration, cloud infra, AI models, document generation, web frameworks, services, and design. Each skill is semantically indexed with 3-tier progressive disclosure.
> 0.6Strong match: load skill body0.4–0.6Review description, may be relevant< 0.4No match: skip this skillThresholds validated against a 120-query eval dataset (90 positive × 30 skills + 30 true negatives). Run make calibrate to recalibrate.
test-writer
pytest, Jest, Go test suites with full edge case coverage and mocking patterns.
claude-api
Anthropic SDK: tool use, streaming, vision, prompt caching, extended thinking, batch.
mcp-server-builder
FastMCP (Python) or TypeScript SDK -- tools, resources, prompts.
docker-containerization
Production Dockerfiles, multi-stage builds, Docker Compose, security hardening.
nextjs-best-practices
App Router -- RSC, async params, image/font optimization, self-hosting.
stripe-integration
Checkout Sessions, webhooks, subscriptions, Connect (Accounts v2), security checklist.
Running in
minutes.
Three paths to a running server. Cloudflare free tier for production. Docker for local-only or air-gapped use.
One command does everything: checks prerequisites, creates .env, installs deps, seeds Qdrant, pushes secrets, deploys the Worker.
.\scripts\setup.ps1
bash scripts/setup.sh
make setup
The wizard prompts for credentials at the right moment. Your server is live once it completes.
All platforms use the same SSE transport config. Replace the URL with your deployed Worker URL, or use the hosted instance above to try immediately without deploying.
{
"mcpServers": {
"skill-mcp": {
"transport": "streamable-http",
"url": "https://skill-mcp.jigneshponamwar.workers.dev/mcp"
}
}
}Replace the URL with your deployed Worker URL. Use the hosted instance to try without deploying.
Drop a platform instruction file into your project root. It teaches the agent the 3-tier skill workflow: when to call skills_find_relevant, how to interpret scores, and when to load tier-3 resources.
cp master-skill/platforms/claude-code/CLAUDE.md ./CLAUDE.mdcp master-skill/platforms/cursor/.cursorrules ./.cursorrulescp master-skill/platforms/windsurf/.windsurfrules ./.windsurfrulescp master-skill/platforms/cline/.clinerules ./.clinerulescp -r master-skill/platforms/copilot/.github ./.githubcp master-skill/platforms/codex/AGENTS.md ./AGENTS.mdcp -r master-skill/platforms/antigravity/.agents ./.agentscp master-skill/platforms/aider/CONVENTIONS.md ./CONVENTIONS.mdAfter copying, edit the file and replace the placeholder Worker URL with your actual deployed URL.