A tool from the drawer that still runs
Every now and then I open my old dotfiles and find something that has quietly survived every machine migration. This time it was AI Shell, a small CLI from Builder.io that I started using more than two years ago. I installed it, almost forgot about it, and the funny thing is: it still works today. No drama, no rewrite, no breaking change that killed it. You type, it answers.
If you have never seen it, the idea is almost embarrassingly simple. You describe what you want in plain English, and it gives you back the shell command that does it. The repo tagline is exactly that honest: a CLI that converts natural language to shell commands. You can find it here: github.com/BuilderIO/ai-shell.
What it actually does
The install is the usual one liner:
npm install -g @builder.io/ai-shell, then point it at an OpenAI key withai config set OPENAI_KEY=...
After that you just talk to it. You write something like ai list all log files larger than 10 megabytes and it proposes a command. You can run it, revise it with another sentence, or cancel. It has a chat mode for when you want to ask instead of execute, a silent flag to skip the explanation, and it understands prompts in more than fifteen languages. Under the hood it leans on OpenAI, originally inspired by the GitHub Copilot X CLI, but open and free to use. It first showed up in April 2023, which in this field feels like a previous geological era.
Why this matters: it was vibecoding before vibecoding
Here is my real take, and the reason I am writing this at all. Before the agents, before Cursor, before Claude Code, before anyone said the word vibecoding out loud, this little CLI was already doing the core move. The term itself did not even exist yet: Andrej Karpathy only coined vibecoding in February 2025, almost two years later. You state intent in natural language. You get an executable result back. That is the entire loop. Say what you want, get the thing.
We tend to tell the story of AI tooling as if it arrived fully formed in 2024 and 2025. It did not. The shape was visible much earlier, just smaller and humbler. AI Shell did not write your whole app and it did not pretend to. It translated one sentence into one command. But that translation, intent to artifact, is exactly the seed that everything else grew from. The difference between then and now is scope and confidence, not the fundamental idea.
What it gets right, and what it does not
I want to stay honest, because the anti-hype is the point. AI Shell is not an agent. It does not hold a plan, it does not loop over your codebase, it does not fix its own mistakes. It is a single shot translator with a confirm step. And that confirm step is actually the part I respect most: it shows you the command before anything runs. You stay in the loop. You are still the one who presses enter.
That restraint feels almost old fashioned now, and I mean that as a compliment. A lot of the friction with modern agents is that they do too much before you can look. AI Shell does one thing, shows it, and waits. There is a lesson in there that we keep half forgetting.
The point of looking back
I am not telling you to switch your workflow to a 2023 CLI. I still reach for the heavy tools when the work is heavy. But it is worth remembering that the idea now reshaping how we build was already sitting in a terminal two years ago, in a tool small enough to forget and stable enough to still answer when you call it.
The good ideas usually arrive quietly first. They show up as something modest, something you almost overlook, and only later does everyone agree it was the beginning. AI Shell was one of those. It said what you want, it got you the thing, and it pointed, without making any fuss about it, straight at where all of this went.