Every comparison, with sources.
We ran Atelier and 10 named code-search tools through the identical 14-repo, 7,213-query benchmark -- same corpus, same query kinds, same scoring. 7 of 10 publish some number about themselves; 1 of 10 has ever benchmarked itself against another search tool on the same axis we measure here. None had been scored against each other before this page existed.
Atelier vs. vanilla Claude Code →
Every published suite: SWE-bench, Exploration, Terminal-Bench, Telegraphic Q&ASame model. Same tasks. Same environment. Anthropic doesn't publish Claude Code's own cost/token/turn efficiency numbers against anything -- so this is the head-to-head Anthropic itself has never run.
A system-prompt-only instruction anyone can paste in today -- the one DIY alternative every reader can reproduce themselves.
Bigger average, way less consistent →A Bash-output compression proxy -- not a code-search tool, so not in the MRR table below. The most-starred tool anywhere in this comparison.
Estimate table, never run against a real task →AST-based semantic code search built on CocoIndex, a Rust data-transformation engine; tree-sitter parsing plus embeddings.
Self-benchmarked, never vs. a rival →Tree-sitter-based persistent knowledge graph (SQLite-backed) across 158 languages -- the most-starred tool in this comparison.
Self-benchmarked, never vs. a rival →In-memory, frecency-ranked file and content search with a background watcher -- fast file discovery, not symbol-level code search.
Self-benchmarked, never vs. a rival →LSP-wrapped semantic code toolkit -- symbol-level navigation and refactoring via real language servers, 40+ languages.
Self-benchmarked, never vs. a rival →Line-oriented regex search CLI -- no AST or symbol awareness, the fastest possible answer to a text-match question.
Benchmarks itself, not vs. this →Tree-sitter AST indexing for 10 core languages, with fallback file indexing for 50+ more.
No published numbers at all →Tree-sitter AST pattern matcher for structural search-and-rewrite -- precise if you already know the exact shape of the code you want.
No published numbers at all →Tree-sitter AST symbol retrieval with a compact wire format (MUNCH) -- optimized for token count, not previously measured for match quality.
Self-benchmarked, never vs. a rival →Local SQLite knowledge graph of symbols, call edges, and dependencies, built via tree-sitter and queried over MCP.
Self-benchmarked, never vs. a rival →Tag/definition indexer for editor jump-to-definition -- exact tag lookups only, no free-text or semantic query support.
No published numbers at all →Methodology, full 13-tool table, and every raw run → BENCHMARKS.md and docs.atelier.ws .