playbooks

home / mcp / self-hosted mcp server

Self-hosted MCP Server

Provides an on-premise verification MCP with verify and verify_document tools for fact-checking prompts and documents.

Installation
Add the following to your MCP client configuration file.

Configuration

View docs
{
  "mcpServers": {
    "kluster-ai-verify-mcp": {
      "url": "http://localhost:3001/stream"
    }
  }
}

You can run kluster.ai’s HTTP Streamable MCP self-hosted server locally to add fact-checking and document validation to your prompts. It includes two tools, verify and verify_document, that help you verify user prompts and agent responses against reliable sources and document content.

How to use

Connect an MCP client or integration to your local MCP server to start using the verification tools. You can run the server locally and configure a client to point at the local stream endpoint. Once connected, you can invoke verify to fact-check conversations and verify_document to ensure responses reflect the provided documents.

How to install

Prerequisites you need before installation are: - A kluster.ai account - A kluster.ai API key - Docker or Node.js 18+ - Git - A client to test the MCP (e.g., Claude Desktop)

Step 1: Clone and set up the MCP server locally ``bash git clone https://github.com/kluster-ai/verify-mcp cd verify-mcp ``

Step 2: Deploy the server using Docker (recommended) or Node.js. Replace YOUR_API_KEY with your actual API key.

Docker (recommended):

docker build -t kluster-verify-mcp .
docker run --rm -p 3001:3001 kluster-verify-mcp --api-key YOUR_API_KEY

Node.js:

npm install
npm run build
npm start -- --api-key YOUR_API_KEY

The server starts at http://localhost:3001 with the MCP endpoint at /stream.

Additional configuration and usage notes

Add the MCP server to your client configuration to enable direct communication with the local stream endpoint. You can connect via standard MCP client configuration or via tooling that supports MCP connections.

Configuration

CLI options you can use with the server:

- --api-key <key> kluster.ai API key.

- --base-url <url> kluster.ai base URL.

- --port <port> Server port (default: 3001).

Connecting to Claude Desktop

To connect Claude Desktop to your local MCP server, add the following configuration and restart Claude Desktop for changes to take effect.

{
  "mcpServers": {
    "kluster-verify": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://localhost:3001/stream"
      ]
    }
  }
}

What you can do with the MCP tools

This MCP provides two tools to support your workflow:

- verify — fact-check a user prompt and the agent’s response against reliable sources.

- verify_document — verify that the agent’s response accurately reflects content from a given document.

Troubleshooting and notes

If you encounter connection issues, ensure the local server is running on the expected port and that the client is pointed to http://localhost:3001/stream. Restart the client after any configuration changes.

Available tools

verify

Fact-check a user prompt and the agent's response against reliable sources.

verify_document

Verify that an agent response reflects the content of the referenced document.