// the find
Sentdex/TermGPT
Giving LLMs like GPT-4 the ability to plan and execute terminal commands
A 2023 proof-of-concept that lets GPT-4 plan and execute shell commands in a loop. The user reviews the proposed commands before they run, which is the main safety mechanism. At 412 stars it found an audience, but it's a notebook-era experiment that was abandoned before it matured.
The human-in-the-loop review step before execution is the right instinct — most agentic shell tools at the time just ran commands blindly. The code is small enough to read in 20 minutes and actually understand what's happening. The accompanying notebook explains the reasoning, making it genuinely educational.
Last commit is July 2023 — three years stale, and the agentic-shell space moved fast. There's no sandboxing or rollback; 'user reviews commands' is the entire safety story, which breaks down the moment a user approves a command they didn't fully parse. It's hardwired to GPT-4 via OpenAI with no abstraction, so the 'future work' about open-source models was never done. The project is a single Python file and a notebook — not a library, not a package, not something you'd actually build on.