Lets you run Python code on Google Colab's free T4 or premium L4 GPUs from any MCP client without local hardware. Three tools: colab_execute for inline code snippets, colab_execute_file for local .py scripts, and colab_execute_notebook to run code and download all generated artifacts like model weights or CSVs. First run triggers OAuth2 in your browser, then tokens cache locally. Handles timeouts, captures stdout/stderr per cell, and validates zip extractions to prevent path traversal. Useful when you need to prototype CUDA kernels, fine-tune models, or run torch/TensorFlow workloads from Claude without spinning up cloud instances.
claude mcp add --transport stdio pdwi2020-mcp-server-colab-exec -- uvx mcp-server-colab-exec