Export as MCP Langchain server

Introduction to the MCP Server for a tested AI Optimizer & Toolkit configuration

This document describes how to re-use the configuration tested in the AI Optimizer and expose it as an MCP tool. The MCP tool can be consumed locally by Claude Desktop or deployed as a remote MCP server using the Python/LangChain framework.

This early-stage implementation supports communication via the stdio and sse transports, enabling interaction between the agent dashboard (for example, Claude Desktop) and the exported RAG tool.

Export config

After testing a configuration in the AI Optimizer web interface, navigate to: Settings/Client Settings:

If—and ONLY if—Ollama or OpenAI are selected as providers for both chat and embedding models:

• select Include Sensitive Settings checkbox
• Click Download LangchainMCP to download a zip file containing a complete project template based on the currently selected configuration.
• Extract the archive into a directory referred to in this document as <PROJECT_DIR>.

To run the exported project, follow the steps in the sections below.

Pre-requisites.

The following software is required:

• Node.js: v20.17.0+
• npx/npm: v11.2.0+
• uv: v0.7.10+
• Claude Desktop free

Setup

With uv installed, run the following commands FROM <PROJECT_DIR>:

uv init --python=3.11 --no-workspace
uv venv --python=3.11
source .venv/bin/activate
uv add mcp langchain-core==0.3.52 oracledb~=3.1 langchain-community==0.3.21 langchain-huggingface==0.1.2 langchain-openai==0.3.13 langchain-ollama==0.3.2

Standalone client

A standalone client is included to allow command-line testing without an MCP client.

From <PROJECT_DIR>, run:

uv run rag_base_optimizer_config_direct.py "[YOUR_QUESTION]"

This is useful for validating the configuration before deploying the MCP server.

Run the RAG Tool by a remote MCP server

Open rag_base_optimizer_config_mcp.py and verify the MCP server initialization.

• For a Remote client, ensure the configuration matches the following:

# Initialize FastMCP server
mcp = FastMCP("rag",host="0.0.0.0", port=9090) #Remote client
#mcp = FastMCP("rag") #Local

• Next, verify or update the transport configuration:

    #mcp.run(transport='stdio')
    #mcp.run(transport='sse')
    mcp.run(transport='streamable-http')

• Start the MCP server in a separate shell:

uv run rag_base_optimizer_config_mcp.py

Quick test via MCP “inspector”

• Start the MCP inspector:

npx @modelcontextprotocol/inspector@0.15.0

• Open the generated URL, for example:

http://localhost:6274/?MCP_PROXY_AUTH_TOKEN=1b40988bb02624b74472a9e8634a6d78802ced91c34433bf427cb3533c8fee2c

• Configure the inspector as follows:

  • Transport Type: Streamable HTTP
  • URL: http://localhost:9090/mcp

• Test the exported RAG tool.

Claude Desktop setup

The free version of Claude Desktop does not natively support remote MCP servers. For testing purposes, this limitation can be addressed using the mcp-remote proxy.

• In Claude Desktop, navigate to Settings → Developer → Edit Config. Edit the claude_desktop_config.json to add a reference to the remote MCP server in streamable-http:

  {
  "mcpServers": {
  	...
  	,
  	"rag":{
  		"command": "npx",
  		"args": [
  			"mcp-remote",
  			"http://127.0.0.1:9090/mcp"
  			]
  		}
  	}
  }

• Next, go to Settings → General → Claude Settings → Configure → Profile, and update the fields such as:

  • Full Name
  • What should we call you

  and so on, putting in What personal preferences should Claude consider in responses? the following text:

  #INSTRUCTION:
  Always call the rag_tool tool when the user asks a factual or information-seeking question, even if you think you know the answer.
  Show the rag_tool message as-is, without modification.

  This forces the use of rag_tool for all relevant queries.

• Restart Claude Desktop.

• You may initially see warnings related to the rag_tool configuration. These warnings are expected and do not prevent tool activation.

• When starting a conversation, Claude Desktop will prompt you to authorize the use of the rag tool:

If the question matches content stored in the vector store, the response will be grounded in that knowledge base. Otherwise, the LLM will fall back to its general training or other configured tools.

nvm -list