monty

Why we included this project

Monty is aimed at teams whose agents need to run code the model just wrote, without standing up a container sandbox for every task. It embeds a Rust-based Python interpreter that handles a deliberate subset of the language, with startup in the single-digit microseconds rather than hundreds of milliseconds. Host access is off by default: filesystem, environment variables, and network only come into play through functions the developer explicitly exposes to the agent. Because the interpreter snapshots its state to bytes at external function calls, you can pause a long-running agent run and resume it later from a file or database. It is still experimental and intentionally thin on libraries.