Get it freeChecked October 9, 2026
GIT_REAL: free safety checks for AI agents
Tools
Use it now
Get it free
Workshop Access
In the workshop
GIT_REAL
AI coding assistants leave half-finished work behind: unsaved edits, brand-new files, stray side branches. Then one stops and asks what to do with the mess, and you cannot tell whether it is junk or an hour of work.
GIT_REAL assesses what one exact action, such as deleting new files or resetting to an older version, would put at risk. You get a clear ALLOW or BLOCK with the reasons, for you and for your assistant. It does not run the action, and it does not give anyone permission to run it. It is free and open source.
Who it helps
- People who use Claude Code, Codex or Cursor on real projects and worry about lost work
- People who run several assistants across several copies of a project
- Anyone who wants a written check in front of risky actions
What you can do
- See what is unsaved, unpublished or easy to overlook in a project.
- Ask about one exact action and read the ALLOW or BLOCK with its reasons.
- Watch many projects on one wall, the riskiest first.
- Give your assistant instructions to read the verdict before it acts.
- Turn on an optional check that refuses a save when something looks unsafe.
Good to know
- GIT_REAL assesses. It does not run any action and it does not give permission. A person or an assistant still decides, and an ALLOW for one action is never permission for another.
- It is advice, not a lock. It only helps if your assistant or a check you turn on follows it. It does not search your files for passwords, it is not a backup, and it cannot prove your work reached a server.
- The free edition on GitHub is the public one. The set-up service on this site uses my fuller private edition.
- Setup runs in a terminal window on your own computer, on Linux, macOS or Windows, with one free extra program that its instructions walk you through. Try it on a practice project before real work.
Get started
The lab
- 33 new working tools. One session.
ChatGPT + Codex built 33 new DMR_REMOTE tools - Useful review, inconsistent grades
Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistency - Extra review cost more on both repairs
Does an extra AI review step pay off? A small DeepSeek trial said no