Every coding agent needs to understand your codebase. But does it really need to remember it?
Most code intelligence tools maintain indexes, embeddings, or knowledge graphs, depends on other stuff like sqlite. These need to stay synchronized with your code, or risk giving agents outdated answers.
I built cxj to try a simpler approach: read the current source, answer the question, and forget everything.
What is cxj?
An open-source, stateless Java/Spring code intelligence CLI for coding agents.
Instead of spending tokens exploring dozens of files, your agent gets structured answers in milliseconds.
cxj brief # Understand the entire service cxj endpoints # List HTTP routes cxj trail PaymentService.pay # Follow calls and find tests cxj callers PaymentService.pay # Find callers cxj deps # Get resolved dependency versions It understands Spring-specific concepts, resolves method calls using declared types, and explicitly reports uncertainty instead of silently guessing.
Does it actually help?
I compared cxj against traditional grep + file exploration using the same AI model on a production Spring service.
For six structural questions, both approaches answered correctly:
| Improvement with cxj | |
|---|---|
| Input tokens | 6.4× fewer |
| Agent cost | 5.6× cheaper |
| Completion time | 2.2× faster |
For simple one-line lookups, grep actually won. cxj is most useful when answers span multiple files.
Full benchmark and methodology.
No index. No daemon. No stale cache.
cxj uses tree-sitter to parse the current code on every invocation, builds its knowledge in memory, and exits.
- 561-file production Spring service: ~170 ms
- Spring Framework, 1.5 million lines: 2.4 seconds
No persistent state, no background process, and no network required for queries.
Try it
brew install umuterturk/tap/cxj npx skills add umuterturk/cxj [download] (https://github.com/umuterturk/cxj/releases/latest) Works with Claude Code, Codex, OpenCode, and other skill-compatible agents. macOS and Linux. MIT licensed.
GitHub: https://github.com/umuterturk/cxj
Come contribute!
I'm curious whether other developers find persistent codebase indexes necessary, or whether parsing on demand is already fast enough for most projects.
