I played around with codegraph. I did a quick experiment where I gave claude code a feature to implement, and I had it perform the same task with and without codegraph. This wasn't scientific, so take it all with a grain of salt.

The codegraph implementation used fewer tokens. However, I performed an automated code review on each implementation, and there were more issues with the codegraph version. These weren't major issues, more like it re-implemented a function that already existed, or used styling that wasn't consistent with the rest of the codebase. My theory is that there's useful information an LLM gleans while grepping through code that is missed when it uses codegraph instead.

It's good, and my agent is using it just fine (explicitly told it to prefer it over just grepping randomly). I use it in Swift, Go, and Node projects. If you use worktrees, make sure to run `codegraph init` post-setup to avoid confusing the robot :)