fastroute

Local MCP plugin: when an agent receives a new task, retrieve Top-N recommendations from all locally installed skills (user-level/plugin-level/project-level). It only recommends, never executes.

The indexing method comes from SkillRouter (arXiv 2603.22455): all-field segmented multi-vector (one vector for name+description + one vector for each ~800-character chunk of body text); for a query, cosine similarity is computed against each chunk, and skill scores are max-pooled.

Tools

Related MCP server: skills-mcp-server

Installation (Claude Code)

claude mcp add fastroute -- node /absolute/path/to/fastroute/dist/index.js

Data and Models

• Data directory ~/.fastroute (overridden by FASTROUTE_DATA_DIR): fastroute.db, config.json, models/

• Embedding model: Xenova/paraphrase-multilingual-MiniLM-L12-v2 (q8, ~135 MB), downloaded online when the tool is first invoked; defaults to the ModelScope mirror (FASTROUTE_EMBEDDING_HOST can point back to https://huggingface.co); if a ~/.knowyoubetter/models cache already exists, it is reused directly

• The model is a hard dependency: when offline and no cache is available, the tool returns an error and download instructions; there is no lexical fallback

• config.json can override: chunkSize(800), autoThreshold(0.9), promptThreshold(0.6), topK(5), userSkillsDir, pluginCacheDir, projectSkillsEnabled

• After changing chunkSize, existing indexes are not automatically re-chunked: delete ~/.fastroute/fastroute.db and then call reindex_skills

Development

npm install
npm test            # 单测 + 集成（fake embedder，不联网）
npm run test:real   # 真实 embedding 冒烟（首次下载模型）
npm run eval        # distractor 评测：Hit@1/Hit@5/MRR
npm run build
npx @modelcontextprotocol/inspector node dist/index.js