What CodeGraph says, vs. what it scored.
Local SQLite knowledge graph of symbols, call edges, and dependencies, built via tree-sitter and queried over MCP.
“Surgical context, fewer tool calls, faster answers, 100% local.”
“CodeGraph hands the agent the exact code it needs in one call. It's a pre-built knowledge graph of every symbol, call edge, and dependency in your codebase.”
| Tool | MRR | p95 | p100 |
|---|---|---|---|
| ★ Atelier +semantic (BGE) | 0.727 | 390ms | 1057ms |
| ★ Atelier lexical (default) | 0.676 | 134ms | 319ms |
| CodeGraph | 0.296 | 17ms | 532ms |
CodeGraph's own numbers are WITH-CodeGraph vs. WITHOUT, on the same repos Atelier's own Exploration suite happens to use -- a real result, but a different question from "does it find the right code." On our shared 7,213-query set it's 0.296 MRR, the second-lowest score in the field, with the fastest raw latency of any tool here (17ms p95) -- it answers almost instantly, just not always with the right symbol.
The true story
Every tool in this comparison, CodeGraph included, has been through the exact same 14 repositories and 7,213 query/gold pairs that score Atelier — no cherry-picked queries, no separate corpus. Full methodology, every raw number, and the other 9 tools →