What we're
building next.
Three phases shipping in sequence. The core vision: the boundary between a skill and an MCP server should dissolve. Open a Discussion before starting major work.
Foundation
Core infrastructure shipped and production-ready. Self-hostable on Cloudflare Workers with Qdrant for semantic retrieval.
Core infrastructure shipped and production-ready. Self-hostable on Cloudflare Workers with Qdrant for semantic retrieval.
MCP-native server on Cloudflare Workers
FastMCP + SSE transport via src/worker.py
Semantic retrieval via Qdrant Cloud
384-dim bge-small-en-v1.5 embeddings, cosine similarity
30 production-ready skill procedures
Sourced from Anthropic, Vercel, Stripe, Cloudflare
Three-tier progressive disclosure
Frontmatter, Body, References / Scripts / Assets
CI/CD skill validation
GitHub Actions to block invalid or malicious skill content in PRs
Automated setup scripts
One-command setup for Windows, Linux, and Make
Agent Intelligence
Making tools self-describing, closing workflow gaps, and improving agent compliance with the three-tier lookup protocol.
Making tools self-describing, closing workflow gaps, and improving agent compliance with the three-tier lookup protocol.
Self-describing tool preconditions
Tool descriptions encode preconditions and postconditions to guide agent behavior directly. No external instruction files needed.
skills_list_all() tool
Browse the complete skill catalogue without a search query
skills_suggest_workflow() tool
Multi-step task planning from a single natural-language request
Agent compliance instrumentation
Measure why agents call get_body without first calling skills_find_relevant
Evaluation framework
Benchmark and calibrate score thresholds for skill retrieval decisions
Workflow enforcement
Ensure the find_relevant, get_body lookup sequence is respected in every session
Infrastructure
Modernising transport, decoupling core dependencies, and building a secure community contribution pipeline.
Modernising transport, decoupling core dependencies, and building a secure community contribution pipeline.
Streamable-HTTP transport
Replace SSE with the modern streamable-http MCP transport
Multiple skill registry support
Point the server at more than one Qdrant collection or remote registry
Decoupled embedding model
Swap bge-small for any Workers AI or external embedding endpoint
Enhanced security scanner
Prompt-injection detection at both seed time and query time
Community YAML validation pipeline
Automated syntax checks and duplicate detection for contributed skills
Cross-platform procedure translation
Automatically adapt skill instructions for different agent architectures
Want to
contribute?
Open a Discussion before starting major changes. Read src/worker.py and the master-skill files together to find the highest-impact improvements.