Launching timesfm-mcp: Give your AI Agents Time-Series Forecasting Superpowers

Have you ever tried to make an AI agent predict the future? Language models are brilliant at reasoning, but they notoriously struggle with numeric time-series prediction out of the box.

Today, we're releasing timesfm-mcp, an open-source MCP server that gives agents native access to robust forecasting.

Why use timesfm-mcp?

• Instant Start (No ML bloat): The server defaults to a statistical baseline (seasonal-naive + trend) that is fast, mathematically sound, and installs in milliseconds.
• Google's TimesFM 2.5 Support: If you need the heavy-hitting foundation model, timesfm-mcp lazily loads and wraps Google's new 200M parameter TimesFM 2.5 model. Your agents can automatically predict massive series natively.
• Backtesting Built-in: We've included a backtest tool that calculates MAE and sMAPE, allowing agents to mathematically prove that their forecasting choices are robust before presenting them to users.

Quickstart

With MCP, adding forecasting to your favorite agent (like Claude Code, Cursor, or Claude Desktop) is just one line:

uvx timesfm-mcp

Check out the repo here: https://github.com/ramdhavepreetam/timesfm-mcp