v0.10 · APACHE-2.0 · LOCAL-FIRST · NO API KEY · NO TELEMETRY archex. Local code context for agents. Turns a repository into a ranked, token-budgeted context bundle — symbols, dependencies, graph context, provenance. It returns context, not an answer. THE PIPELINE   repository → context bundle 01 Parse tree-sitter 25 languages 9 full · 15 chunk 02 Index BM25F (lexical) + local vectors + SPLADE 03 Retrieve RRF fusion + rerank + graph expansion 04 Assemble token-budgeted bundle, deduped, dependency-closed MEASURED · 19-TASK HEAD-TO-HEAD · APPLE M1 PRO archex competition 1.0 Recall 0.95 0.32 F1 0.66 0.31 Token eff. 0.76 0.48 COMPLETION-PENALTY TOKENS 922 vs 11,188 tokens for the agent to finish · ≈12× fewer COLD START 0 ms vs 4,721 ms no daemon warm-up required The real baseline isn’t another index — it’s grep. grep gets recall 1.00 and token-efficiency 0.00 — it dumps everything into the prompt. archex hands back only the relevant slice, at a fraction of the tokens. SHIP IT ANYWHERE · ONE DETERMINISTIC CORE CLI · 21 commands MCP · 14 tools Python API Docker slim / full Claude Code skill LangChain retriever LlamaIndex retriever scout: map → fetch graph query + analyze $ uv tool install archex github.com/ Mathews-Tom/archex Deterministic retrieval · runs fully local · same query = same bundle, every time.