RAT MCP Server (Retrieval Augmented Thinking)

What is RAT MCP Server?

A Model Context Protocol (MCP) server that implements RAT’s two-stage reasoning process, combining DeepSeek’s reasoning capabilities with various response models.

How to use RAT MCP Server?

Clone the repository, install dependencies with npm install, create a .env file with API keys and model configuration, then build the server with npm run build. Add it to Cline’s MCP settings with the generate_response tool.

Key features of RAT MCP Server

• Two-stage processing: DeepSeek reasoning then response generation.
• Supports Claude 3.5 Sonnet, DeepSeek Reasoner, and OpenRouter models.
• Maintains conversation history with configurable context size.
• Single tool generate_response with optional reasoning display.
• Context clearing via clearContext parameter.

Use cases of RAT MCP Server

• Use DeepSeek reasoning followed by a high‑quality Claude response.
• Keep conversation context across multiple interactions.
• Debug by inspecting the reasoning process with showReasoning.
• Reset the conversation for a new topic using clearContext.

FAQ from RAT MCP Server

What API keys do I need?

You need a DeepSeek API key and an OpenRouter API key. An Anthropic API key is optional if you use Claude models.

How do I clear the conversation context?

Set clearContext: true in the generate_response tool arguments.

Can I see the reasoning process?

Yes, set showReasoning: true in the tool call to display DeepSeek’s reasoning.

What models can be used for final response?

Claude 3.5 Sonnet (via Anthropic) or any model available on OpenRouter, such as GPT-4 or Gemini.

Is there a context size limit?

Yes, the server supports a configurable context size limit (adjustable via settings).