XERJ

XERJ is a single static Rust binary that runs a local search server. Point autoindex at a folder and it builds typed indices with no schema work. Search is BM25, kNN, and hybrid; the default embedder is lexical and offline, with an optional neural embedder. It speaks Elasticsearch-compatible APIs so existing clients keep working, exposes MCP plus an install that wires Claude Code or Cursor, and includes agent-memory namespaces for durable notes.

Agents still burn tokens on grep and full-file reads. XERJ turns a project — code, docs, logs, PDFs — into ranked passages and go-to-definition hits, so the model works from the right slice. Measured reference-coding runs cut output tokens ~2.7× vs grep at the same solve rate, which is real money and fewer fix loops on private or post-cutoff code.

Install with the one-liner from xerj.org, start the server on loopback, run xerj init to attach MCP and the skill, then xerj autoindex on your repo or a folder of reference projects. Agents search by natural language or symbol, get file:line passages, and can run reference coding — clone close open-source repos, index them, and reuse how they solved a problem before writing new code. xerj gain reports what it actually saved.

Not another vector RAG toy or Mac desktop search. XERJ is agent-native local search built for coding workflows: autoindex, ES-compatible API, MCP, and measured token wins on reference coding. Complementary to PageIndex (tree RAG over long PDFs), DigUp (on-device multimodal Mac search), and codebase-memory-mcp (code knowledge graph) — those solve different retrieval shapes; this one kills wasteful grep in the agent loop.

• Public demo available
• API / SDK surface
• MCP server / client support

• PageIndex ↗
• DigUp ↗
• codebase-memory-mcp ↗