DEV Community

Devair
Devair

Posted on AI-assisted

I wrapped Google TimesFM-3 as an MCP server — you still cannot use the weights in production

#python #mcp #opensource #machinelearning

Google released TimesFM-3 on 31 August 2026. I shipped an MCP server for it the next day so Claude Desktop, Claude Code, and Cursor can call the model as a tool.

Repo: https://github.com/thenameisdevair/timesfm3-mcp

What the agent actually gets

The tool is forecast(history, horizon).

You pass a chronological list of numbers. You get:

  • forecast — median / point prediction
  • quantiles — the nine official TimesFM-3 heads, q10 through q90
  • a license note on every response

TimesFM-3 already produced those quantile heads. Most thin wrappers throw them away. This one does not.

The license is the product constraint

Piece License
This MCP wrapper Apache-2.0
TimesFM-3 weights (google/timesfm-3.0-pytorch) non-commercial, non-production

You can research, evaluate, and run agent experiments. You cannot put this checkpoint behind a paid API, a customer deliverable, or a live demand planner. Google’s commercial path is BigQuery / AlloyDB AI.FORECAST.

If a README hides that, it is not a serious integration.

Install

Weights are gated on Hugging Face. Accept the model terms first.

git clone https://github.com/thenameisdevair/timesfm3-mcp.git
cd timesfm3-mcp
python3 -m venv venv
source venv/bin/activate
pip install -r local/requirements.txt
pip install git+https://github.com/google-research/timesfm.git
huggingface-cli login
cd local && python client.py
Enter fullscreen mode Exit fullscreen mode

Point Claude or Cursor at local/server.py using the venv Python. Config snippets are in the README.

What this release is not

No multivariate targets yet. No past/future covariates. No commercial backend. Those are next. This cut exists so “TimesFM-3 MCP” has a public, honest surface while the launch is still warm.

If the tool shape is wrong for your agent, feel free to open an issue on the repo.

Top comments (0)