What problem does it solve?
It eliminates the friction of launching ML experiments by automatically deploying, preparing, and running jobs across local GPUs, remote SSH servers, Vast.ai instances, and Modal serverless GPUs.
Core Features & Use Cases
- Environment-aware deployment: Detects whether to run locally, via SSH, on Vast.ai, or on Modal by reading project
CLAUDE.md, then follows the correct lifecycle.
- GPU pre-flight validation: Checks GPU availability before starting experiments (e.g., via
nvidia-smi and memory-usage heuristics) to avoid failed or wasted runs.
- Code syncing and job launch: Syncs code efficiently (rsync or git), optionally injects Weights & Biases logging, and starts the experiment with captured logs.
- Cost-conscious execution: Supports Vast.ai auto-destroy after completion and Modal’s scale-to-zero behavior to reduce idle charges.
- Use Case: You want to train a new model variant and need reliable execution plus logged results on the best available GPU provider without manually rewriting deployment steps each time.
Quick Start
Ask to run the experiment by providing your experiment description and requesting deployment on the configured environment in your project CLAUDE.md.
