What jCodeMunch says, vs. what it scored.
Tree-sitter AST symbol retrieval with a compact wire format (MUNCH) -- optimized for token count, not previously measured for match quality.
“The leading, most token-efficient MCP server for precise GitHub source code retrieval via tree-sitter AST parsing... cut AI token costs 95%+ on code exploration.”
| Tool | MRR | p95 | p100 |
|---|---|---|---|
| ★ Atelier +semantic (BGE) | 0.727 | 390ms | 1057ms |
| ★ Atelier lexical (default) | 0.676 | 134ms | 319ms |
| jCodeMunch | 0.299 | 214ms | 4189ms |
The most rigorous self-benchmark in this whole field -- real repos (Express, FastAPI, Gin), a published methodology file, a 95% token-reduction number that holds up. What it doesn't measure is whether the retrieved symbol is the right one: on our matched accuracy run it's 0.299 MRR, near the bottom of the field. Efficient retrieval of the wrong symbol is still the wrong symbol.
The true story
Every tool in this comparison, jCodeMunch 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 →