toto-mcp

# Toto MCP Server An MCP (Model Context Protocol) server for the Toto time series forecasting model by Datadog. ## Installation ### Step 1: Clone this repository ```bash git clone git@github.com:Ankcorn/toto-mcp.git cd toto-mcp ``` ### Step 2: Clone the Toto repository The Toto model source code needs to be available locally: ```bash git clone https://github.com/datadog/toto.git ``` ### Step 3: Install dependencies This project uses `uv` for Python package management: ```bash uv sync ``` ### Step 4: Test the installation ```bash uv run python toto_server.py --help ``` ## Usage To add this MCP server to Claude, add the following configuration to your MCP settings: ```json { "mcp-server-toto": { "command": "uv", "args": [ "run", "--directory", "/path/to/toto-mcp", "python", "toto_server.py" ] } } ``` ## Available Tools ### `forecast_timeseries` Forecast time series data using the Toto model. **Parameters:** - `series_data`: List of time series, each as a list of float values (channels x timesteps) - `prediction_length`: Number of future timesteps to predict (default: 336) - `num_samples`: Number of samples for probabilistic forecasting (default: 256) - `samples_per_batch`: Control memory usage during inference (default: 256) - `time_interval_seconds`: Time interval between observations in seconds (default: 900 = 15 minutes) **Returns:** Dictionary containing forecast results with median, quantiles, and samples. ### `get_model_info` Get information about the Toto model and system capabilities. ## Available Resources ### `toto://model-info` Get Toto model information as a resource. ### `toto://usage-example` Get usage example for the Toto forecasting tool. ## Requirements - Python 3.8+ - PyTorch - CUDA (optional, for GPU acceleration) - Internet connection (for downloading the pre-trained model on first use) ## Notes - The Toto model will be downloaded automatically on first use - GPU acceleration is automatically enabled if CUDA is available - Recommended input length is 4096 timesteps for best results - The model supports multivariate time series forecasting