KumoRFM MCP Server

KumoRFM MCP Server – Empowering AI Assistants with Relational Foundation Model Intelligence

The KumoRFM MCP server bridges the gap between advanced relational data analytics and conversational AI. It exposes a pre‑trained Relational Foundation Model (RFM) that can generate predictions, fill missing values, and forecast temporal trends directly from multi‑table datasets—all without the need for any training. By exposing this capability through MCP, developers can embed powerful data‑driven reasoning into agents, chatbots, and other AI workflows with minimal friction.

At its core, the server translates natural language into a Predictive Query Language (PQL) query that runs against the KumoRFM model. The MCP interface allows an AI assistant to request a graph construction from raw CSV or Parquet files, ask high‑level questions (“What’s the projected sales trend for next quarter?”), and receive structured, model‑generated answers. The server also provides tools to visualize the underlying graph, inspect predictions, and evaluate results—all accessible via standard MCP tool calls.

Key capabilities include:

• Graph construction from relational data – automatically converts tabular files into a heterogeneous graph representation, preserving schema relationships.
• Natural‑language to PQL translation – lets agents formulate queries in plain English, which are then converted into precise predictive commands.
• Training‑free predictions – the underlying RFM delivers accurate imputation, forecasting, and anomaly detection without any user‑supplied training data.
• Interactive visualization – developers can inspect the graph and prediction outputs directly from within an agent’s toolset.

Typical use cases span finance, e‑commerce, and operations. For example, a sales forecasting agent can ingest transactional tables, ask the MCP to forecast revenue for the next month, and receive a concise answer along with confidence intervals. In healthcare, an assistant could predict patient readmission probabilities from multi‑table electronic health records without training a separate model.

Integration is straightforward: any MCP‑compatible client (Claude Desktop, CrewAI, LangGraph, etc.) can launch the server as a subprocess or via an MCP bundle. The agent then calls the server’s tools, receives structured responses, and can chain them with other reasoning steps. This tight coupling enables end‑to‑end workflows where conversational agents not only answer questions but also generate actionable data insights on demand.

The KumoRFM MCP server stands out for its zero‑training predictive power and native graph‑based reasoning, making it a compelling addition to any developer’s AI toolkit who needs robust, relational analytics without the overhead of model training or complex data pipelines.