sourcekeep

SourceKeep

Git was built in 2005. Software isn’t built that way anymore.

Git was made for the Linux kernel. People emailed patches to a mailing list. Almost twenty years later, the way we build has changed. Most code is now written by AI agents. Teams need privacy that “public repo or private repo” can’t give them. And we just put up with the daily friction of branches, commits, and worktrees. We stopped asking why.

SourceKeep asks why. And it’s being built in the open.

The one thing git got wrong

Most problems with git come from one choice. Git copies the whole repo to every clone, and turns it into real files on your disk.

That one choice is why a few things are hard.

• Privacy is all or nothing. A clone is everything. So you can’t keep one file like .env, one branch, one package, or one in-progress security fix private inside a shared repo.
• You have to commit to share. The commit is the thing that gets copied around. So history is something you manage by hand instead of something that just happens while you work.
• Worktrees are a pain. A second working copy is a second copy you babysit. It can’t even bring your node_modules or your .env.

Fix the foundation and the rest follows.

The idea

The repo is a database, not a folder. Your working copy is a view of it, not the real thing.

One real copy lives on the server. It keeps the good parts of git, like storing data by its content. It drops the rest, like copying everything to every machine. On top of that, three things git can’t do.

1. Private by path

Open source shouldn’t mean all of your code, public, all of the time.

• Check in your .env. Share it with your team. Keep it out of the public repo.
• Keep a branch or a whole package private while main stays public.
• Fix a security bug in private. Show the diff when you’re ready, not the moment you open the PR.
• Share a PR with anyone you choose. The whole public for open feedback, or three teammates for sensitive work. You decide who sees it.

GitHub can’t do this. A git clone needs the whole repo, so it’s all or nothing. That’s why this has to be a clean break.

2. Built for the agents writing your code

Most code is written by agents now. They shouldn’t have to drive a 2005 command-line tool and parse its text.

• A real API to read, write, and search your code. No filesystem needed.
• A record on every change of why it was made, which model made it, and how it was checked. When AI writes most of the code, that’s the new code review.

3. A worktree that actually works

Start a fresh, throwaway work area from any commit in milliseconds. Unlike git worktree, it brings everything you need to run.

• Your source, from that commit.
• Your dependencies. node_modules is already there. No npm install.
• Your config. .env is already there, because SourceKeep tracks it for you, in private.

Start ten from the same commit. Each one is isolated. Let an agent work in each. Keep the one whose tests pass. Throw the rest away. No worktrees to manage. No containers to set up and tear down. No waiting.

A clean break from git

This isn’t git with a compatibility layer. Git’s model is the problem. Being git-compatible would just bring back the limits we want to remove. A real git clone has to copy the whole repo, and that’s what makes private files in a repo impossible. Being git-native would mean we couldn’t do the one thing SourceKeep is for.

So we keep the good ideas, drop the rest, and give you a one-way path to bring your code in.

Open, and yours

• MIT licensed. Run your own copy for free.
• Your data is yours. Open, documented export. No lock-in.
• A hosted version is coming for teams who’d rather not run it themselves.

This lives on GitHub until SourceKeep can host itself. Moving it onto SourceKeep is the first milestone.

This is early, and I want your take

Right now SourceKeep is a direction and a thing being built in the open. It’s not finished. I’ve written software for about 27 years, and I think its foundation is wrong for how we build today.

This is the thinking so far. If it speaks to you, or if you think I’m wrong, I want to hear it.

Follow along or reach me on X at @jtwebman, or use the contact form at jtwebman.com.

This site is open source. Improve this page.