ExpertPack MCP server

ep-mcp

<!-- mcp-name: io.github.brianhearn/ep-mcp -->

Expertise-as-a-Service over the Model Context Protocol.

An MCP server that turns any ExpertPack into a live, queryable knowledge service. Any MCP-compatible agent (Claude Desktop, Cursor, Windsurf, Claude.ai, custom hosts) can connect and retrieve high-quality, provenance-rich domain expertise via standardized tools.

🌐 expertpack.ai · 📦 ExpertPack Framework · 📖 Architecture

97.6% retrieval hit rate / 0 full misses on a 22-question benchmark against a production v4.1 atomic-conceptual pack (658 chunks). Prior v4.0 baseline on the same pack and eval set: 65.1%.

Features

• Schema-aware hybrid retrieval: BM25 + vector search (sqlite-vec), metadata boosting, MMR re-ranking, adaptive threshold filtering, length penalty
• requires: expansion (schema v4.1 atomic-conceptual): When a top-K atom's frontmatter declares requires: [B, C], the referenced atoms are auto-appended to results as additional SearchResults flagged requiresexpanded=True. Directional (A→B does not imply B→A), transitive up to 2 hops, capped at 3 atoms and 3,500 cumulative tokens, appended after the maxresults slice so it never displaces top-K.
• Intent-aware routing: Automatic query classification (ENTITY/HOW/WHY/WHEN/GENERAL) adjusts vector/BM25 fusion weights per query — no LLM required
• Query embedding cache: SQLite-backed cache keyed by (model, dimension, sha256(query)) — warm query latency drops from ~4s (live Gemini API) to ~1ms
• Deep graph traversal: Optional multi-hop BFS expansion over knowledge graph edges with per-hop score decay and drift prevention
• Reserved-slot graph merge: Configurable protected slots in the final top-K reserved for graph-expanded neighbors so related files surface even against dominant core results
• BM25 fallback mode: Embedding provider outage degrades gracefully to lexical-only search (coverage + density + path + length scoring) instead of a hard error
• Cross-encoder reranker: Optional second-pass precision reranker (disabled by default, ~70ms warm latency when enabled)
• Provenance-first: Every result includes id, contenthash, verifiedat, source_file
• Reconstruct mode (RFC-003): Opt-in reconstruct=true / reconstruct: true enriches results with fragmentid, linerange, originalmarkdown, span-level contenthash, excerpt, stale, and byte offsets so agents can prove exactly what source text grounded an answer
• EP-native chunking: Files are treated as atomic retrieval units (split at ##/### only for oversized content). Split chunks keep the original span as content/spanhash, carry a computed linerange, and record a heading_path (0.7.0+)
• Heading paths (0.7.0+): every hit carries headingpath (e.g. ["Routing", "Setup", "Ports"]) so an H3 hit keeps its H2 parent; epread(path, section=..., neighbors=true) reads just that section plus its previous/next siblings

Shortened. The full README is on GitHub.