Shared Knowledge

You can run a shared knowledge MCP server that powers several AI assistants with fast, context-aware information retrieval. It uses retrieval augmented generation to search a knowledge base across multiple vector stores, enabling consistent access for different tools while keeping you in control of how data is stored and searched.

How to use

Set up a local MCP server that you can connect to from your AI assistants. Start the server in stdio mode so it runs as a local process and expose its capabilities to clients via the MCP protocol. You’ll point your MCP-enabled client at the local node process and supply environment variables to tailor how the knowledge base is loaded, how vectors are stored, and how results are filtered.

Typical usage involves starting the MCP server once, then querying it from your client when you need knowledge base lookups. You can configure the server to use a particular vector store, limit retrieval results, and control chunking behavior to balance speed and precision. If you work with different assistants or toolchains, you can reuse the same MCP endpoint so all tools share a unified knowledge base.

How to install

Prerequisites: you need Node.js installed on your system. Ensure you have a working OpenAI API key if you intend to generate embeddings via OpenAI’s models.

Install and run the MCP server locally using the following steps. Copy and adapt the commands to your environment.

# 1) Clone the MCP server repository (replace with your actual path if you have a local copy)
git clone https://example.com/shared-knowledge-mcp.git
cd shared-knowledge-mcp

# 2) Install dependencies
npm install

# 3) Start the server (example for development)
npm run dev

Configuration and connection methods

This server is designed to run as a local process that your MCP-enabled clients can connect to. The following configuration example shows how to start the server as a stdio-based MCP service, with environment variables controlling its behavior.

{
  "mcpServers": {
    "sharedkb": {
      "command": "node",
      "args": ["/path/to/shared-knowledge-mcp/dist/index.js"],
      "env": {
        "KNOWLEDGE_BASE_PATH": "/path/to/your/rules",
        "OPENAI_API_KEY": "your-openai-api-key",
        "SIMILARITY_THRESHOLD": "0.7",
        "CHUNK_SIZE": "1000",
        "CHUNK_OVERLAP": "200",
        "VECTOR_STORE_TYPE": "hnswlib",
        "VECTOR_STORE_CONFIG": "{}"
      }
    }
  }
}

Environment variables you will configure

Use these environment variables to control how the MCP server loads data, searches, and stores vectors. Provide your own values or placeholders as needed.

KNOWLEDGE_BASE_PATH=\/path\/to\/your\/docs
OPENAI_API_KEY=your-openai-api-key
SIMILARITY_THRESHOLD=0.7
CHUNK_SIZE=1000
CHUNK_OVERLAP=200
VECTOR_STORE_TYPE=hnswlib
VECTOR_STORE_CONFIG={}

Tools and capabilities

The MCP server exposes a retrieval tool that searches the knowledge base to fetch relevant documents, summarize results, extract keywords, and explain relevance. This enables you to incorporate precise, contextual information into responses from your AI assistants.

Available tools

rag_search

Search the knowledge base using a retrieval augmented generation approach, returning matching documents with optional summaries, keywords, and relevance explanations.