跳到正文
DEV Community· kurojack-jp·· 16 小時前AI 評分51

Sora Commander v0.5.0 更新,簡化 Windows 安裝與更新流程

My Codex 5-Hour Limit Hit 0%. ChatGPT Chat Still Created the File (Sora Commander v0.5.0)

AI 導讀

開發者 kurojack-jp 發佈 Sora Commander v0.5.0,並示範在 Codex 的 5 小時限額顯示剩餘 0% 時,透過普通 ChatGPT 聊天使用這個 Windows MCP 伺服器建立檔案。

正文

Recently my Codex 5-hour limit showed 0% left. I still had work to finish, so I opened an ordinary ChatGPT chat and asked it to create a file on my Windows PC through Sora Commander, my open-source MCP server. A small approval window popped up on my desktop. I read it, clicked Approve, and the file appeared in Explorer.

Here is that run, 26 seconds, recorded on my PC on 2026-11-21 (captions are in Japanese):

On ChatGPT Plus with the Codex 5-hour limit at 0%, ChatGPT asks Sora Commander to create a file, I approve it on my PC, and the file appears

An honest caveat first: this shows that this request still ran. It does not show that tool calls never count toward any limit. ChatGPT's own message limits still apply, and plan terms can change at any time. What it does show is that the reading, running and checking happen in the ChatGPT chat I already use, so that work isn't spent from a separate coding agent's limit.

That run was on the build I've just released: v0.5.0.

What Sora Commander is (30-second version)

If you missed the first post: Sora Commander is a free, open-source (AGPL-3.0) MCP server for Windows. It connects to ChatGPT through OpenAI's Secure MCP Tunnel. ChatGPT can read and search files in folders I allow, run commands and report their output, and create or edit supported text files. Commands that aren't plainly safe, and every file create or edit, open an approval window on my PC and run only if I approve them there. There is no remote approval.

What's new in v0.5.0

This release is mostly about getting started and staying up to date with less friction.

Setup, key and update helpers

  • tools\setup.cmd creates .venv, installs the packages, copies the example tunnel settings if they're missing, runs the doctor check and lists what's left to do by hand. --check only reports.
  • tools\set-tunnel-key.cmd stores the tunnel key with hidden input, in a file only your Windows user can open. --check checks its form without showing it.
  • tools\update.cmd previews incoming releases with --check, then fast-forwards a clean main.
  • tools\install-broker-autostart.cmd (optional) starts Broker mode and then the tunnel when you sign in to Windows.

No PowerShell needed

The launchers and helpers now use only .cmd files and Python. They keep working if Windows PowerShell 5.1 becomes optional, and PowerShell 7 isn't required either. The prerequisites are now Windows 11, Python 3.13 and Git for Windows in C:\Program Files\Git.

Shorter tool help, same tools

Shared approval rules moved into the server instructions, so each tool description is shorter. Tool names and inputs are unchanged.

  • With the included launchers, ChatGPT sees 54 tools.
  • The optional read-only Muse bridge lists 36 tools, or 37 with --inbox. The six command and wait tools it could never run are gone from its list. (I wrote about the bridge in this post. It's read-only: commands are refused and no approval window opens from it. Muse is a Meta product; this project is not affiliated with or endorsed by Meta.)

Fixes I'm happy about

  • Git tools now show Japanese and other non-ASCII file names as they are, instead of \343\203... escapes. As a Japanese developer, this one was personal.
  • bridge_digest and tool_catalog report 0.5.0 instead of UNKNOWN.
  • A key check error keeps the launcher window open so you can actually read it.
  • Owner ACL errors are now a single [HOLD] line naming the file and the reason.

Quick start

Before anything else, check your ChatGPT plan. Custom MCP connections are web-only, and OpenAI's help page lists full MCP for Business and Enterprise/Edu. I use it on Plus with developer mode, including approved commands, as of November 2026, but Plus isn't listed there and that can change.

Then, roughly (the full steps are in docs/INSTALL.md):

git clone https://github.com/kurojack-jp/SoraCommander.git C:\SoraCommander
cd C:\SoraCommander
tools\setup.cmd
  1. Download tunnel-client from OpenAI's Secure MCP Tunnels guide (it isn't bundled) and put it in tools\tunnel-client\<version>\.
  2. Create a tunnel in the OpenAI platform settings, then run tools\set-tunnel-key.cmd to store its key.
  3. Fill in the three values in %USERPROFILE%\.config\tunnel-client\sora-next.yaml, and list the folders you want to allow.
  4. Double-click tools\start-broker-mode.cmd and keep both windows open.
  5. Connect ChatGPT as OpenAI's help page describes, then in a new chat ask for ping, status and tool_catalog. tool_catalog should report 54 tools and version 0.5.0.

Updating from v0.4.0: v0.4.0 has no update helper, so update once by hand: stop the tunnel, the Broker and the Muse bridge, then in a clean checkout on main run git pull --ff-only and tools\setup.cmd --check. From then on, tools\update.cmd does it. Afterwards, restart everything and refresh the tools in ChatGPT.

Safety: what the approval window does and doesn't do

I'd rather be precise here than reassuring:

  • Commands that aren't plainly safe, and file edits, need a click at the PC. If the window can't open, the session is locked, or records can't be written, it fails closed. That also means you have to be at the PC to approve.
  • Reads need no window. ChatGPT can read your allowed folders while you're signed in, including when Broker mode is started at logon.
  • An approved command runs with your own Windows rights and can do anything your account can. The window and folder limits reduce risk; they are not a sandbox. Read each window before you click.
  • In Broker mode, a command whose outcome is unknown isn't repeated, even after a restart, until you clear it by hand.
  • It's a developer preview. status reports remote_ready: false and LOCAL_DEMO_ONLY, the code isn't signed, and there's no installer. v0.5.0 was checked on my own setup and in a fresh Windows Sandbox (where setup.cmd and the launch checks passed from an ASCII and a Japanese checkout path, without starting a tunnel). That isn't production acceptance, and there's no warranty.

Try it

Issues are welcome. For security problems, please follow SECURITY.md instead of opening a public issue.

The repo and its install docs are free. If you'd rather follow a step-by-step path with screenshots and the common errors explained, there's also a paid companion guide: Sora Commander — The Complete Beginner's Guide (Gumroad).

Sora Commander is an independent project, not made, endorsed or supported by OpenAI.

Sora Commander started as "please stop making me copy-paste between the chat and my terminal." v0.5.0 doesn't change the limits; it just makes the setup shorter, so there's less standing between you and the chat you already have open.

來源:DEV Community · dev.to