What it does

Kaguya is a ChatGPT plugin that exposes a local API for file operations and script execution. You drop it in a Docker container, point ChatGPT at localhost:3000, and the model can list files, read them, edit via search-and-replace, or run Python/JS/bash commands. Your files live in a FILES folder; everything outside Kaguya’s own directory is off-limits.

The interesting bit

The README is refreshingly honest about the friction. It warns you to keep files under 100 lines, explicitly instruct the model to use search-and-replace instead of overwriting whole files, and even suggests restarting the conversation if ChatGPT starts hallucinating. This is a tool built by someone who has actually watched an LLM confidently destroy working code.

Key highlights

• Sandboxed to a single directory—no wandering into ~/.ssh
• File ops include granular search-and-replace, append, rename, and directory management
• Shell command execution via executeCommand (with the usual “please cd first” caveat)
• Dockerized setup; one script to run
• Custom instructions provided to keep the LLM from being overly eager

Caveats

• Requires OpenAI’s plugin devtools, which are waitlist-gated and may not be generally available
• Project is no longer active; author points to safeclaw as the successor
• VS Code extension and open alternative mentioned but not delivered

Verdict

Worth a look if you’re studying early ChatGPT plugin patterns or want a reference for LLM-local-system bridges. Skip it if you need something maintained—follow the redirect to safeclaw instead.