Source · repo
XERJ: local search engine and MCP server for agents
Summary
XERJ is a single Rust binary that indexes code, docs, logs, CSV/JSONL, and PDFs on your machine and serves keyword (BM25), vector, and hybrid search through an Elasticsearch-compatible API and a stdio MCP server. It is Apache-2.0 for the engine, with about 19K stars and over 4,000 commits (read 2026-11-16; latest release candidate v1.0.0-rc.101). The date above is the read date; the repo is not new but was rising in the GitHub trends report.
Key points
- Install:
curl -fsSL https://xerj.org/get | sh(read the script first), a Nix flake, orcargo build --release -p xerj-serverinengine/. - Run and index:
xerj --insecure --data-dir ~/xerj-data &starts a local node (--insecureturns off TLS and API keys: single-user, localhost only), thenxerj autoindex ~/my-project. - Use from Claude Code:
claude mcp add xerj -- /absolute/path/to/xerj mcp(or anmcpServersentry in.mcp.json); setXERJ_URLif the node is not onhttp://localhost:9200, and pass the key fromadmin.keyviaXERJ_AUTHwhen auth is on.xerj mcponly proxies to a running node. Tools includexerj_searchandxerj_def. - Good for: giving an agent fast search over a large repo plus its logs and docs without reading whole files into context.
- Caveats: the token-saving and accuracy figures are the project's own and some benchmark scripts are unpublished; the optional XERJ Query-1 answer model shipped in release binaries since rc.101 is proprietary; the README is heavily promotional.
Why it matters
A self-hosted alternative to cloud doc and code search MCP servers: one binary, local data, and a familiar Elasticsearch API for other tools.