内容
# TimesFM-3 MCP Server
MCP server that exposes Google's **TimesFM-3** zero-shot time-series foundation model to AI agents (Claude Desktop, Claude Code, Cursor, and any MCP client).
Pass one series or several related series, plus optional past / known-future covariates. Get a median forecast plus nine quantile bands (`q10`–`q90`) per series. Optional `start` + `freq` (or a history timestamp list) labels each forecast step with a date.
**Repo:** https://github.com/thenameisdevair/timesfm3-mcp
## License — read this first
| Piece | License | What you can do |
| --- | --- | --- |
| This MCP wrapper | Apache-2.0 | Use, modify, ship the *server code* |
| TimesFM-3 **weights** (`google/timesfm-3.0-pytorch`) | `timesfm-non-commercial-license-v1.0` | Research, evaluation, non-production experiments only |
Do **not** put this server on a paid product, customer workflow, or production planner while it loads the public TimesFM-3 checkpoint. Google's commercial path is BigQuery / AlloyDB `AI.FORECAST`.
The `forecast` tool repeats this warning in every response.
## What the agent gets
Tool: `forecast`
Univariate (also accepted as `history` for older clients):
```json
{
"series": [[10.5, 12.1, 14.8, 15.2, 18.0]],
"horizon": 5
}
```
Joint multivariate with a known-future promo flag (`future_covariates` length is `T + horizon`):
```json
{
"series": [
[10, 11, 13, 12, 14, 16, 15, 17],
[20, 21, 22, 24, 23, 26, 25, 28]
],
"series_ids": ["sku_a", "sku_b"],
"future_covariates": [[0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1]],
"horizon": 3
}
```
Response (multivariate):
```json
{
"status": "success",
"model": "google/timesfm-3.0-pytorch",
"mode": "multivariate",
"n_series": 2,
"context_length": 8,
"horizon": 3,
"series": [
{
"id": "sku_a",
"forecast": [17.2, 17.8, 18.1],
"quantiles": {
"q10": [15.0, 15.4, 15.7],
"q50": [17.2, 17.8, 18.1],
"q90": [20.1, 21.0, 21.6]
}
}
],
"quantile_levels": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9],
"license": "TimesFM-3 pretrained weights are non-commercial / non-production..."
}
```
Numbers above are illustrative. Univariate calls also include top-level `forecast` and `quantiles` so older clients keep working.
Shape rules (same as TimesFM-3):
- All target rows have length `T`
- `past_covariates` each have length `T`
- `future_covariates` each have length `T + horizon`
`forecast_demand(history, horizon)` remains a deprecated alias for a single series.
Optional calendar (does not change the model output, only labels it):
```json
{
"history": [10, 11, 13, 12, 14, 16, 15, 17],
"horizon": 3,
"start": "2026-08-25",
"freq": "D"
}
```
```json
{
"status": "success",
"freq": "D",
"history_end": "2026-09-01",
"timestamps": ["2026-09-02", "2026-09-03", "2026-09-04"],
"forecast": [17.7, 18.6, 19.7]
}
```
`freq` is `H`, `D`, `W`, or `M`. `start` is the first history timestamp. If you pass a `timestamps` list instead, it must be length `T` and strictly regular — gaps are an error, not filled. Omit all of these and the response stays array-only.
## Local setup
Weights are gated on Hugging Face. Accept the model terms, then log in.
```bash
git clone https://github.com/thenameisdevair/timesfm3-mcp.git
cd timesfm3-mcp
python3 -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r local/requirements.txt
pip install git+https://github.com/google-research/timesfm.git
huggingface-cli login
```
Shape/validation tests (no weight download):
```bash
python -m unittest tests.test_forecasting -v
```
Smoke-test inference (loads the 330M checkpoint; first run downloads weights):
```bash
cd local
python client.py
python client_calendar.py
```
Start the stdio server the same way an MCP client will:
```bash
cd local
python server.py
```
RAM: plan for roughly 16GB+ on CPU. A GPU is optional; the server uses CUDA when PyTorch sees it.
## Connect an agent
### Claude Desktop
Add to `claude_desktop_config.json`:
```json
{
"mcpServers": {
"timesfm3": {
"command": "python",
"args": ["/ABS/PATH/timesfm3-mcp/local/server.py"]
}
}
}
```
Use the interpreter from the venv that has `fastmcp`, `torch`, and `timesfm` installed. Example:
```json
"command": "/ABS/PATH/timesfm3-mcp/venv/bin/python"
```
### Claude Code
```bash
claude mcp add timesfm3 -- /ABS/PATH/timesfm3-mcp/venv/bin/python /ABS/PATH/timesfm3-mcp/local/server.py
```
### Cursor
Add to `~/.cursor/mcp.json`:
```json
{
"mcpServers": {
"timesfm3": {
"command": "/ABS/PATH/timesfm3-mcp/venv/bin/python",
"args": ["/ABS/PATH/timesfm3-mcp/local/server.py"]
}
}
}
```
Then ask: *Jointly forecast sku_a and sku_b for 8 steps, with this promo flag as a future covariate, and give q10 / q50 / q90.*
## Cloud (Hugging Face Space)
`cloud/` is a Docker image that serves the same tool over HTTP/SSE on port 7860.
1. Create a **Docker** Space.
2. Set the build context to the `cloud/` folder (it includes `forecasting.py`), or copy `cloud/Dockerfile`, `cloud/app.py`, `cloud/forecasting.py`, and `cloud/requirements.txt` to the Space root.
3. Add Space secret `HF_TOKEN` (a Hugging Face token that can download `google/timesfm-3.0-pytorch`).
4. Keep the Space marked research / non-commercial. The weights license does not allow a paid hosted forecast API.
## Status
Implemented now:
- Univariate and joint multivariate `forecast` with point + 9 quantiles
- Past-only and past-future covariates
- Optional `start`/`freq` or regular history timestamps on the response
- Local stdio + cloud SSE
- Explicit license banner on the tool and in this README
Not in this release:
- Commercial inference backend (TimesFM 2.5 or BigQuery)
MCP 市场