Recommendation Engine

The recommendation stage is implemented in recommender/engine.py and recommender/llm_reasoning.py.

Inputs

• analyzed task (TaskDefinition)
• optional preferred provider
• optional Tavily API key

Data Sources

• web search via Tavily (WebSearchProvider)
• MCP registry data from:
  • official MCP registry (registry.modelcontextprotocol.io)
  • Smithery fallback (registry.smithery.ai)
• skills catalogs from GitHub:
  • anthropics/skills
  • github/awesome-copilot

Parallel Fetch

Recommendation fetches in parallel:

• web results
• relevant MCP servers (filtered)
• full skills catalog

LLM Output Contract

LLM must return JSON with:

• llm provider/model/reason
• framework
• mcp_servers
• skills
• prompt_strategy
• estimated_tokens
• estimated_cost

Skill Filtering Rules

Returned skills are filtered against the fetched skill catalog.

• unknown skills are dropped
• ambiguous names without source are dropped
• canonical output is "name (source)"