Shell-GPT - AI-Assisted Command Shell

Shell-GPT

My Electron shell experiment for turning GPT-generated command suggestions into bounded Docker-backed execution.

I built Shell-GPT as an experiment in a safer AI-assisted command workflow. The app wraps GPT command generation in a desktop shell and uses Docker as an execution boundary so generated commands can be tried in a more controlled environment.

Summary

Shell-GPT is an Electron-based shell experiment that explores how an AI assistant can move from suggesting commands to helping run them inside a bounded environment. The point is not to blindly trust generated shell output; it is to make the boundary around that output clearer.

Problem

AI command suggestions are useful, but shell commands are high-impact. A wrong command can delete files, leak data, mutate state, or leave a system in a confusing condition. I wanted to prototype a workflow where AI-generated commands could be inspected and executed with a stronger separation from the host machine.

My Role

I built the application concept around an Electron shell, a TypeScript interface, OpenAI-backed command assistance, and Docker-based execution. The project let me explore the interaction design around human review, command generation, and sandboxed execution.

Stack

The project uses TypeScript, Electron, React, OpenAI APIs, Docker, CSS, and a small amount of JavaScript. The public repository is TypeScript-first and includes Dockerfile support for the execution side of the experiment.

Architecture

The desktop app provides the shell-like interface. The assistant layer produces command suggestions from user intent. Docker provides a container boundary for trying commands without treating the host machine as the default execution target.

Interesting Technical Decisions

The execution boundary is part of the product idea, not an afterthought.
The interface frames AI output as something to inspect and run deliberately rather than as an invisible automation step.
Electron made sense for the prototype because the workflow needed a local desktop surface, process integration, and a controlled UI.

Status

Experimental prototype. The repository is public and useful as a record of the AI shell concept, but the README still needs to be replaced with project-specific documentation.

Links

Repository and related topic hub: AI-assisted development.

Install

bash

# npm (global)
npm i -g shell-gpt

# pnpm (global)
pnpm add -g shell-gpt

# bun (global)
bun add -g shell-gpt

# from source
git clone https://github.com/Kalaghni/shell-gpt
cd shell-gpt && pnpm i && pnpm build && pnpm link --global

Configuration

.env

# OpenAI
export OPENAI_API_KEY=sk-...

Shell-GPT

My Electron shell experiment for turning GPT-generated command suggestions into bounded Docker-backed execution.

Problem

Interesting Technical Decisions

The execution boundary is part of the product idea, not an afterthought.

The interface frames AI output as something to inspect and run deliberately rather than as an invisible automation step.

Electron made sense for the prototype because the workflow needed a local desktop surface, process integration, and a controlled UI.

# npm (global)

npm i -g shell-gpt

# pnpm (global)

pnpm add -g shell-gpt

# bun (global)

bun add -g shell-gpt

# from source

git clone https://github.com/Kalaghni/shell-gpt

cd shell-gpt && pnpm i && pnpm build && pnpm link --global