pnegahdar / nano

What is nano?

Nano is a minimal coding agent that runs entirely in Python without external dependencies. You point it at a task—`./nano.py "fix the tests"`—and it loops with an LLM, running shell commands on your machine after showing you what it plans to do. It defaults to OpenAI's Responses API but accepts any model via environment variable. The agent reads your project's CLAUDE.md or README.md for context, discovers skill files from common agent directories, and keeps working until the job completes or hits the 200-command limit.

Why is it gaining traction?

The hook is simplicity: most agent frameworks ship as sprawling codebases you must understand before trusting. Nano is 200 lines you can audit over lunch. It turns the "black box" concern on its head—you can read every line. The human-in-the-loop approval model balances automation with safety, showing each command before execution. Session resume means you can hand off mid-task or continue later without local state, since OpenAI stores the conversation. The interactive REPL enables exploratory sessions, while one-shot mode lets you queue-and-forget simpler tasks.

Who should use this?

Developers who want to understand what coding agents actually do under the hood will get the most value—Nano is transparent where others are opaque. Backend engineers tired of verbose YAML configs for simple automation might prefer this stripped-down approach. Teams evaluating agent frameworks could use it as a baseline comparison: if GPT-5.5 plus a shell loop handles your use case, you may not need Kubernetes, Docker, or a plugin system. Anyone cautious about fully autonomous agents will appreciate that approvals are on by default.

Verdict

Nano earns attention for its philosophy, but the credibility score of 0.75% signals early-stage software with minimal community vetting. At 39 stars, it has not yet proven itself at scale. The pure stdlib approach and MIT license make it trivial to fork and own, which mitigates risk—but you should test thoroughly before trusting it with anything beyond throwaway tasks. Try it as a learning tool or a lightweight alternative to heavyweight agent frameworks, but do not mistake small code for battle-tested code.