I published a small tool yesterday: Sora Commander, a free, open-source MCP server that lets ChatGPT read files and run commands on my own Windows PC — right from an ordinary chat.
In the chat, I just ask: "run the tests in this folder and tell me the result." ChatGPT runs the command on my PC, checks the result, and reports back in the same conversation. No separate app, no terminal window to open.
Why I built it
The motivation was simple. There is a paid tool that does this (Remote Desktop Commander), and I wanted a free one for myself. So I built it.
The original goal was just "run a command from chat." But as I built it, the priorities shifted toward making it genuinely usable:
- If an AI runs commands on my PC, I want to see what it is about to run — with my own eyes, before it runs.
- If the connection drops or the PC restarts, I want to recover the result instead of wondering what happened.
- A command whose outcome is unknown must never be re-run by accident.
The approval window is the whole point
Before anything risky runs, an approval window opens on my own PC screen, showing the exact command. Only what I click "approve" on actually runs. There is deliberately no remote approval: to approve, you have to be sitting at the PC.
File access is scoped to folders I list by hand in a file. The chat side cannot rewrite that list.
Honest status: v0.1 developer preview
I want to be straight about where this is:
- Still a v0.1 developer preview. No production acceptance yet (the tool itself reports
remote_ready: false,LOCAL_DEMO_ONLY). - Setup needs Python 3.13 and manual steps. There is no installer yet.
- Paths containing Japanese characters are untested — ASCII paths like
C:\SoraCommanderare recommended. - No code signing.
This is a tool for people who can judge for themselves what a command does. What it can and cannot do is written honestly in the README.
A timely detail: it runs from Instant chat
One thing I forgot to say in my first announcement: no special mode is needed. It works in the ChatGPT chat you already use, including the default Instant mode.
That matters more from December, when OpenAI shares one usage allowance across ChatGPT and Codex on Plus / Pro / Business plans — while its help page says "Instant chat will not draw from your usage." Working from Instant chat means the PC work doesn't eat the allowance you'd spend on Codex. (OpenAI only states this for chat itself; I make no claim about tool calls.)
Try it
- GitHub: https://github.com/kurojack-jp/SoraCommander
- Setup instructions are in the repo (Japanese docs included).
If something breaks, an Issue helps. Security issues only — please use the repo's private vulnerability reporting, not a public Issue.
If you want the guided walkthrough
The setup instructions in the repo are free. If you'd rather follow the whole setup as one guided path — real screenshots at every step, the common errors and what to do about them, and the security settings explained — I also wrote it up as a complete beginner's guide: Sora Commander — The Complete Beginner's Guide (Gumroad, $15)
I have a previous version of this project, by the way. I made the security so rigid that it stopped working, so I threw it away and rebuilt it from scratch. The lesson stuck: security mechanisms are not better the heavier they are. The right question for a personal tool is "how much of this do I actually need?"

Top comments (6)
Approval only at the PC, with no remote approve, is a strong choice, and never re-running a command whose outcome is unknown is the detail most tools skip. Scoping folders in a file the chat side cannot rewrite also closes an obvious hole. Since you have been using it yourself, has there been a command you denied in the window, and what in the request made you stop?
Thanks for reading so closely — those three points are exactly the ones I cared about most.
Honest answer: I have never actually denied one. The setup is arranged so the chat side doesn't propose dangerous commands in the first place — the allowed folders are narrow, and I ask for work in small, plain steps, so what reaches the window is already scoped to what I intended.
The window is the backstop, not the first line of defense. Its everyday value is that I see the exact command before it runs — and knowing the window is there is part of what keeps the requests simple. I'd rather the safety be boring than dramatic.
Boring safety is the right goal, and seeing the exact command before it runs is worth it even if you never deny one. The one thing a window that never says no can hide is a habit of clicking allow without reading.
Have you thought about sending yourself a command you should deny now and then, just to check the window still gets read?
That's a sharp point — a window you always approve can quietly become a window you stop reading. I haven't run deliberate deny-drills. What keeps me honest so far is asking for work in small, plain steps, so each window shows one short command I can actually read before clicking. Your drill idea is worth trying: send myself a command I should refuse — something outside the allowed folders, or a destructive one — and check I catch it. One related change since: in v0.1.7, closing the window (X) no longer counts as a reject — it means "no decision", so nothing runs and ChatGPT waits instead of treating it as a no. That helps against a misclick, but it doesn't replace actually reading. I'll try the drill and report back if it catches me out.
Treating the X as "no decision" rather than a reject is a nice detail, since a misclick shouldn't read as a deliberate no. Small, plain steps is the right habit too, because a short command is one you actually read. If the drill catches you even once, that's the most useful result it can give.
That's the test of it — if the drill never catches me, it isn't testing anything. I'll report back here if it does.