timesfm-mcp

Local MCP server for GPU-backed TimesFM 2.5 forecasting.

This package exposes Google's TimesFM model to MCP clients over stdio, sse, or streamable-http. It is intended for local or trusted-network serving where agents need zero-shot time-series forecasts, prediction intervals, covariate forecasting, CSV forecasting, and interval-based anomaly scoring.

What It Provides

• guidance: agent-facing usage guide for safe TimesFM forecasting.
• health: package, CUDA, system, and model-state report with a CUDA matmul probe.
• estimate_memory: rough dataset memory estimate before loading large jobs.
• warmup: lazy-load and compile TimesFM on the GPU before live requests.
• forecast_values: forecast in-memory numeric series.
• forecast_csv: forecast numeric columns in a local CSV and write CSV or JSON.
• forecast_with_covariates_values: TimesFM 2.5 XReg forecasts with known future covariates.
• detect_anomalies: compare future actuals against q20/q80 and q10/q90 forecast bands.
• timesfm://forecasting/guide: MCP resource with the same operational guidance.
• timesfm_forecasting_guide: MCP prompt for agents before planning a forecast.

The model is a process singleton. It is loaded on first warmup or forecast call and remains in memory until the MCP server process exits. If a request changes model settings such as max_context, max_horizon, batch_size, or infer_is_positive, the server reloads the model with the new settings.

When To Use

Use this server for zero-shot univariate time-series forecasting:

• Sales, demand, revenue, traffic, inventory, and capacity planning.
• Sensor readings, vitals, load, weather, prices, and measurements.
• Probabilistic forecasts where q10 through q90 prediction bands matter.
• Known-future-covariate forecasts, such as price, promotion, holiday, weather,

store attributes, product family, or region effects.

• Forecast-vs-actual anomaly review using prediction intervals.

Do not use it for classification, clustering, causal interpretation, coefficient analysis, general tabular prediction, or model fine-tuning. Fine-tuning is a training workflow and is intentionally not exposed by this inference MCP server.

GPU Setup With uv

For RTX 5090 and other new NVIDIA GPUs, install a PyTorch wheel that supports the GPU architecture before installing this package. CUDA 12.8 wheels are the recommended starting point for this machine class.

git clone https://github.com/chokukil/timesfm-mcp.git
cd timesfm-mcp

uv venv .venv-gpu --python 3.10
source .venv-gpu/bin/activate

uv pip install --upgrade --reinstall \
  torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/cu128

uv pip install -e ".[gpu]"
uv pip check

.[gpu] installs the TimesFM torch, XReg, and Flax-related extras, including einshape. If you only need standard torch forecasting without XReg/Flax dependencies, install .[torch]. If you need XReg but not Flax, install .[xreg].

Validate CUDA before starting MCP:

uv run --python .venv-gpu/bin/python python - <<'PY'
import torch
print(torch.__version__)
print(torch.cuda.is_available())
print(torch.cuda.get_device_name(0))
x = torch.randn((512, 512), device="cuda")
y = x @ x
torch.cuda.synchronize()
print("cuda matmul ok")
PY

Start The Server

Start with a conservative batch size. Do not set CUDA_VISIBLE_DEVICES= unless you intentionally want to hide the GPU.

export TIMESFM_BATCH_SIZE=64
export PYTHONNOUSERSITE=1

timesfm-mcp --transport sse --host 0.0.0.0 --port 8765

The SSE endpoint will be:

http://<host>:8765/sse

For local stdio clients:

timesfm-mcp --transport stdio

For streamable HTTP:

timesfm-mcp --transport streamable-http --host 0.0.0.0 --port 8765

Environment Variables

Set infer_is_positive=false per request, or TIMESFM_INFER_IS_POSITIVE=false for the process, when the metric can go below zero: returns, residuals, PnL, temperature anomalies, z-scores, or signed deltas.

Agent Workflow

1. Read guidance or the timesfm://forecasting/guide resource.
2. Call health. If cuda_probe.passed is false, fix PyTorch/CUDA before loading.
3. Call estimate_memory for large workloads.
4. Call warmup once if latency matters.
5. Use the forecasting tool that matches the input shape.
6. Interpret quantiles carefully: index 0 is mean, then q10 through q90.

q10 and q90 form the central 80 percent prediction interval. q20 and q80 form the central 60 percent interval.

Tool Examples

Forecast Values

{
  "inputs": [[10, 12, 11, 13, 15, 16, 18, 19, 21, 22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43, 45, 46, 48, 49, 51, 52, 54, 55]],
  "horizon": 7,
  "names": ["sales"],
  "infer_is_positive": true
}

Forecast With Covariates

Dynamic covariates must have length len(input_series) + horizon for each series. The tail values are known future covariates.

{
  "inputs": [[100, 101, 103, 105, 104, 106, 108, 109, 111, 113, 112, 114, 116, 118, 119, 120, 122, 124, 123, 125, 127, 129, 130, 132, 133, 135, 137, 138, 140, 141, 143, 145]],
  "horizon": 4,
  "dynamic_numerical_covariates": {
    "price": [[9.9, 9.9, 9.8, 9.8, 9.7, 9.7, 9.7, 9.6, 9.6, 9.6, 9.5, 9.5, 9.5, 9.5, 9.4, 9.4, 9.4, 9.3, 9.3, 9.3, 9.2, 9.2, 9.2, 9.1, 9.1, 9.1, 9.0, 9.0, 9.0, 8.9, 8.9, 8.9, 8.8, 8.8, 8.8, 8.8]]
  },
  "static_categorical_covariates": {
    "region": ["seoul"]
  },
  "xreg_mode": "xreg + timesfm"
}

Detect Anomalies

{
  "inputs": [[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32]],
  "actuals": [[33, 60]],
  "horizon": 2,
  "names": ["metric"]
}

Severity rules:

• normal: actual is inside q20 to q80.
• warning: actual is outside q20 to q80 but inside q10 to q90.
• critical: actual is outside q10 to q90.

Security

The SSE and streamable HTTP transports do not add authentication by themselves. Bind to 127.0.0.1 for local-only use. Bind to 0.0.0.0 only on a trusted network or behind your own authentication, firewall, or reverse proxy.

Development

uv pip install -e ".[gpu,dev]"
pytest -q
ruff check .

License

Apache-2.0. This repository wraps TimesFM and depends on the upstream timesfm Python package and model weights.