xerj

XERJ is a local search engine packaged as a single static Rust binary with no JVM. Running xerj autoindex on a folder sniffs every file and creates typed indices for code, docs, logs, PDFs, SQLite and CSV, with code indexed by symbol and line number through tree-sitter. Queries go through an Elasticsearch-compatible API with BM25, kNN vector and hybrid search, so existing ES clients work unchanged.

Its headline use is reference coding: the agent clones open-source repos close to the task, indexes them, and retrieves how they solved the problem before writing code. The project's own case study (8 tasks, 16 runs per arm) reports 2.7x fewer output tokens than a grep-driven agent at the same 16/16 solve rate, and it states openly that the gain disappears on popular libraries the model already knows.

xerj init wires it into Claude Code or Cursor as an MCP server plus skill, and a /_memory API gives agents long-term memory. The default embedder is lexical and offline; neural embeddings are opt-in (about 90 MB). The quick-start runs with --insecure, meaning no TLS or auth, which is only appropriate on a single-user dev machine. License is Apache-2.0.