timesfm-mcp 0.1.6

MCP server for Google's TimesFM 2.5 foundation model — give any AI agent zero-config time-series forecasting.

timesfm-mcp

MCP server for Google's TimesFM 2.5 — give any AI agent zero-config time-series forecasting.

Plug TimesFM 2.5, Google's 200M-parameter foundation model for time-series, directly into Claude Code, Claude Desktop, Cursor, or any MCP client. The agent calls forecast, gets point predictions + uncertainty bands + a trend/seasonality summary, and writes the explanation itself.

No ML configuration. No data pipelines. One line to run.

Chart generated with the statistical baseline. See "Enable TimesFM 2.5" below to use the full neural model.

Quickstart (30 seconds)

uvx timesfm-mcp        # runs over stdio for local agents

Add to your Claude Desktop / Claude Code / Cursor config:

{
  "mcpServers": {
    "forecast": { "command": "uvx", "args": ["timesfm-mcp"] }
  }
}

Then ask your agent: "Forecast the next 6 months from this revenue data and tell me what to expect."

Enable TimesFM 2.5 (optional)

System requirements: ≥ 16 GB RAM · ~800 MB disk (model weights, downloaded on first use) · PyTorch

Not sure? Skip this — uvx timesfm-mcp already works great on any machine.

pip install "timesfm-mcp[timesfm]"

The TimesFM 2.5 source is bundled inside this package (Apache-2.0, Google LLC) — no separate git clone needed. The server auto-detects it and upgrades automatically; no config change required.

Two backends, zero config

Backend	When active	System requirement	Install
Statistical baseline	Always — default	Any machine	uvx timesfm-mcp
TimesFM 2.5 (Google)	When installed	≥ 16 GB RAM + ~800 MB disk	pip install "timesfm-mcp[timesfm]"

Start with the baseline. It runs on any machine, installs in seconds, and delivers production-ready forecasts. Upgrade to TimesFM only if you need the neural model's extra accuracy and have the RAM for it.

Tools

Tool	What it does
forecast	Forecast a single series with optional uncertainty bands
list_backends	Report which engine is active (timesfm / baseline)
backtest	Hold out the last N points — compare TimesFM vs baseline MAE/sMAPE

Documentation

Full docs in the docs/ folder:

• Getting Started — installation and first forecast
• Client Setup — Claude Desktop, Claude Code, Cursor configs
• Tool Reference — full parameter docs
• Cookbook — SaaS MRR, e-commerce demand, traffic, cloud spend
• How It Works — the math and model

Migrating from forecast-mcp

timesfm-mcp is the renamed continuation of forecast-mcp. Update your install:

pip install timesfm-mcp      # replaces: pip install forecast-mcp
uvx timesfm-mcp              # replaces: uvx forecast-mcp

Update your agent config: change "args": ["forecast-mcp"] → "args": ["timesfm-mcp"].

License

Apache-2.0