You give the agent a CSV and ask a question in plain English. It writes a few lines of pandas, runs that code inside a real Firecracker microVM, reads the actual output, and answers.
It was built as a learning spike to evaluate Eve, Vercel’s agent framework, and to find out whether the whole development lifecycle (model, sandbox and tests) could run on an ordinary laptop with a 4 GB GPU, without sending a single token to the cloud.

