Skillogy

GraphRAG-based skill router for Claude Code. Routes every prompt through a Neo4j knowledge graph of your SKILL.md files so the right skill triggers based on intent — not on whether your wording happens to match a description.

Install

/plugin marketplace add PurpleCHOIms/Skillogy
/plugin install skillogy@skillogy

Prerequisites: Node.js ≥ 20 and Docker (for the Neo4j container). Everything else is automatic — the SessionStart hook starts Neo4j and indexes ~/.claude/skills + project .claude/skills in the background. Until indexing finishes, the hook gracefully passes through with a one-time status message.

Benchmark

29 testbed skills, 29 natural-language queries, same claude --print. Only variable: whether Skillogy's UserPromptSubmit hook is active. Detection criterion: Claude Code emits the correct Skill({skill: "..."}) tool call within 60 s.

| Model | Native | With Skillogy | Improvement | |---|---|---|---| | Claude Haiku 4.5 | 20.7 % | 62.1 % | +41.4 pp | | Claude Sonnet 4.6 | 48.3 % | 100.0 % | +51.7 pp | | Claude Opus 4.7 | 69.0 % | 96.6 % | +27.6 pp |

Hook p95 latency < 10 s. Numbers are from the original Python implementation under legacy-python/bench/; the TypeScript port preserves the routing pipeline byte-for-byte.

License

MIT — see LICENSE.