nano.py
A coding agent implemented in under 200 lines of dependency-free Python, arguing that current models make elaborate agent frameworks unnecessary.
nano.py is a minimal coding agent, deliberately kept under 200 lines with zero external dependencies, built to demonstrate that a capable enough model needs little more than a loop, a shell tool, and human approval to be useful. It reads CLAUDE.md, AGENT.md, AGENTS.md, and README.md automatically for repository context, discovers skill files from the same locations real agent harnesses use, and runs a loop of up to 200 steps that continues until the task is done.
Every command the model wants to run is shown to the user with a short description before it executes, with options to approve one command, approve all remaining commands, or deny and stop; nothing runs without that approval by default. It supports an interactive REPL, a one-shot non-interactive mode, and resuming a previous session, defaults to GPT-5.5, and can use any other model through an environment variable. It intentionally does not install packages, require a config file, need a container, or support plugins.