Ever stare at a perfectly functional codebase and have absolutely no earthly idea how you got there? That’s the subtle, maddening terror of working with AI code assistants like Claude Code CLI. You start a session, prompt your way through a complex refactor, accept the changes, and then… silence. An hour later, you’re left with a working repo, but the memory of which specific edit broke what, or why you even went down that particular rabbit hole, has vanished into the digital ether. Cursor has checkpoints. Claude Code CLI? Not so much.
That’s the gaping hole one developer, clearly as bugged by this as I am, decided to fill. And bless their heart, they actually built it.
Why Does This Matter for Developers?
The fundamental design of many AI code CLIs, including Claude Code, is “fire-and-forget.” You prompt, it edits files, you accept. The conversation rolls on, a relentless march forward. But what if you need to backtrack? What if you want to ask, “What exactly did Claude do in the last 30 minutes?” Or, more practically, “Let me see a diff of my code from four prompts ago.” The out-of-the-box answer is a resounding, deafening silence. Sure, you could commit to Git after every single prompt, but let’s be honest, who does that? And the commit messages would be pure gibberish, barely qualifying as documentation.
This new tool, dubbed ccheckpoints, is a tiny daemon that cleverly hooks into Claude Code’s natural lifecycle. Think of it as an auto-save on steroids, but smarter. It triggers a snapshot of your entire workspace before each message submission. When a session finally wraps up, it neatly closes out that checkpoint group. A background server humming away on port 9271 keeps all this state information, and then there’s a local web dashboard at localhost. This dashboard is your new best friend, letting you browse through sessions, jump back to specific checkpoints, and even view file-level diffs. It’s basically Git, but auto-committed with every AI interaction, neatly scoped per session, and blessedly free of those embarrassing, auto-generated commit messages.
The Inner Workings (No Magic Here)
It’s built with TypeScript on Node, which is hardly surprising. For storage, it’s using SQLite. And this is key: it’s 100% local. No cloud. No authentication hoops to jump through. No elaborate privacy narrative to craft. Just… it runs on your machine. It use Claude Code’s native hook system instead of resorting to the far less elegant method of polling. This means the hooks fire precisely when they need to, leading to a cleaner implementation. The web UI? It’s served directly from the daemon itself. Install it, run it, and you’re off to the races.
An interesting technical choice: the creator debated between storing diffs versus full file snapshots. Diffs are typically smaller, but they quickly become a headache when you’re dealing with file renames, deletions, or even binary files. Full snapshots chew up more disk space, but the benefit, especially for a tool that lives solely on your local machine, is enormous: they’re trivial to navigate. Disk is cheap. Ease of use, especially when you’re already stressed by AI-generated code, is priceless. So, ccheckpoints stores the file content for each checkpoint and computes the diffs on demand within the UI.
npm install -g ccheckpoints
ccheckpoints
Once installed and running, you just use Claude Code like you normally would. Those checkpoints? They’ll start appearing in the dashboard automatically.
The cli is fire-and-forget by design — it edits files, you accept, the conversation rolls on. if you want to ask “what did claude actually do in the last 30 minutes?” or “let me diff against where i was four prompts ago”, there’s nothing in the box.
That’s the problem statement in a nutshell. And this solution, while small, feels like a genuinely useful patch for a growing frustration among developers wrestling with these powerful, yet sometimes forgetful, AI coding partners. It’s not a revolution, but it’s a darn good incremental improvement that addresses a very real pain point. Who is making money here? Right now, nobody but the developer themselves, which, in the world of open source, is often the best endorsement you can ask for.
Where to Find It
The code is available on GitHub: github.com/p32929/ccheckpoints. The creator is actively soliciting feedback—especially edge cases where the hooks might miss something. Stars and forks are welcome if you find it useful, which, if you use Claude Code CLI regularly, you probably will.
