Your assistant,with a memory.
A graph of your repository that the assistant you already use can query instead of grepping. Hosted, or on your own machine.
Retrieval by grep.
Your assistant is good at writing code and has no idea what the code around it does. So it reads: whole files, one after another, working out the architecture again from scratch every session, and paying for it every time.
The questions that matter are the ones grep cannot phrase. What calls this. What breaks if I move it. Which of these two functions is the one actually in the request path. Those are questions about connections, and a search over text has none.
Three things you get.
Hosted or on your own machine, the same graph underneath.
It stops rediscovering
The architecture is read once at parse time, not re-derived from file reads every session.
It answers the hard one
What breaks if I change this becomes a traversal with a path back, not a grep and a hope.
It costs nothing to start
Free plan on the hosted product, no card. Or run the open-source engine on your machine, Apache 2.0, no account.
It will not write your code and it takes no actions. It is a map your assistant reads before it edits, which is a smaller claim than most tools make and an easier one to check.
It is also not a search engine. For finding the passage that reads most like a sentence, an embedding index is the better tool and this does not try to beat it.
Two ways in.
Start free on the hosted product, or run the engine yourself in about five minutes.