Plugin MCP local: cuando el agente recibe una nueva tarea, busca las Top-N recomendaciones entre todas las skills instaladas localmente (nivel de usuario/plugin/proyecto). Solo recomienda, no ejecuta.
El método de indexación proviene de SkillRouter (arXiv 2603.22455): multi-vector segmentado all-field (name+description como un vector + cada fragmento de ~800 caracteres del cuerpo como un vector); la consulta calcula la similitud coseno para cada fragmento y la puntuación de la skill se obtiene con max-pool.
Herramienta | Uso |
| Recupera las skills más relevantes antes de comenzar la tarea; devuelve score, tier (auto/prompt/skip, valor de referencia), nombre de invocación y ruta. |
| Reconstruye el índice de forma incremental después de instalar/eliminar skills (mtime diff), sin necesidad de reiniciar la sesión. |
Related MCP server: skills-mcp-server
claude mcp add fastroute -- node /absolute/path/to/fastroute/dist/index.jsDirectorio de datos ~/.fastroute (se puede sobrescribir con FASTROUTE_DATA_DIR): fastroute.db, config.json, models/
Modelo de embedding: Xenova/paraphrase-multilingual-MiniLM-L12-v2 (q8, ~135 MB), se descarga de Internet la primera vez que se invoca la herramienta, por defecto desde el espejo ModelScope (FASTROUTE_EMBEDDING_HOST puede volver a apuntar a https://huggingface.co); si ya existe la caché de ~/.knowyoubetter/models, se reutiliza directamente.
El modelo es una dependencia estricta: si no hay conexión ni caché, la herramienta devuelve un error y una guía de descarga; no hay respaldo léxico.
config.json se puede sobrescribir: chunkSize (800), autoThreshold (0.9), promptThreshold (0.6), topK (5), userSkillsDir, pluginCacheDir, projectSkillsEnabled
Al modificar chunkSize, los índices existentes no se vuelven a segmentar automáticamente: hay que eliminar ~/.fastroute/fastroute.db y luego llamar a reindex_skills.
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.jsThis server cannot be deployed
Search and discover Agent Skills from the skills.sh registry. Powered by HAPI MCP server.
MCP Server for an Agent Task Marketplace
The CustomGPT.ai MCP server is a fully managed, RAG-powered endpoint that connects large language models with private knowledge bases and external data sources. It provides tools for retrieval-augmented generation queries (send_message), data ingestion (upload_file), and source listing, enabling AI agents to query private documents like PDFs with high accuracy and real-time citations.
Fastio gives your team and AI agents a workspace for each project, bringing source files, comments and new work together. Connect a compatible MCP client to find files, answer questions from project material, and save results where the team can review and use them. \ Sign in through Fastio and choose the workspaces the connection can reach. Read Only connections expose read tools; Read & Write connections also allow changes. \ • **Files and folders:** browse, read, upload, and organize files, with version history and trash. • **Search:** find files by keyword; search file content and meaning when Deep Indexing is enabled and files have been processed. • **Metadata:** read, set, and search metadata fields, including extracted fields where available. • **Shares:** create and manage shares for sending or collecting files. • **Comments and members:** read and add comments, and manage workspace membership within your permissions. • **Activity:** search workspace events to see what changed. \ Hosted by Fastio, with no server installation required. \ • **Claude and most MCP clients:** https://mcp.fast.io/mcp/tools • **ChatGPT:** https://mcp.fast.io/mcp/operations • **ChatGPT and Codex:** Fastio plugin • **Claude Code, Cursor and similar coding agents:** https://mcp.fast.io/mcp/code • **Setup and client-specific capabilities:** https://mcp.fast.io/docs \ Paid plans. Monthly plans include a trial of up to 30 days. A card is required, and billing begins when the trial ends or trial credits run out, whichever comes first. No free plan.
