How big of a codebase do I need to justify using this tool?

[chkiss]

How big of a codebase do I need to justify using this tool? I have a small CLI-based game under 500kb, but I've found that Claude can burn through 50%+ of its session tokens just thinking and reading files on a clean session. Not sure if I need something like this tool or if I have other problems.

[tirth8205]

Honest sizing guidance, now also in docs/FAQ.md:

• Under ~200 files: usually net overhead. Our own formal token-efficiency benchmark reports ratios below 1.0 for small commits — documented, not hidden. Plain grep/agentic search wins there.
• The win scales with corpus size and question shape. Single-symbol lookups ("where is X defined") don't need a graph at any size. Multi-hop questions — blast radius of a diff, callers-of-callers, which flows a change touches, what's untested — are where the graph pays off, increasingly so as the repo grows (our pinned-repo evals run 60 → 27k+ files; the best case, 528x, is the largest corpus).
• Review workflows specifically (risk-scored detect-changes on PRs) benefit even on mid-size repos because the alternative is re-reading every touched file plus its dependents.

Rule of thumb: if your repo fits comfortably in your agent's context window, you don't need this tool.