What is XERJ?
XERJ is a local search engine aimed at AI coding agents. Point xerj autoindex at a folder and it works out what each file is, parses code with tree-sitter so functions arrive with their symbols and line numbers, and builds one index per dataset it finds, across code, Markdown, PDFs, CSV, JSON, SQLite, Office files, email and logs. Agents then query it for passages instead of reading whole files, through 20 to 23 MCP tools or the CLI. It also offers vector and hybrid search, a /_memory API for agent memory, and Corpus Hub, a catalogue of pre-built indexes of open-source projects.
Under the hood it implements the Elasticsearch REST API, so Kibana and the official clients connect unchanged; its conformance suite passes 1,444 of 1,454 cases. The project began on 30 June 2026, is written in Rust under Apache 2.0, and is still a release candidate: v1.0.0-rc.100 shipped on 27 October 2026. It has about 18,800 GitHub stars.
Is XERJ worth it?
XERJ is the most ambitious of the tools that give a coding agent search instead of grep, and it asks for more trust than most. The engineering is serious. A single static binary indexes a 25,000-file tree in about two and a half minutes on the project's example, speaks enough of the Elasticsearch API to pass 1,444 of 1,454 conformance cases, and, on the maintainers' head-to-head with Elasticsearch 8.13.4, records 55 wins, 26 ties and 4 losses, with 1.72 times the ingest rate. All four losses are read latency under heavy writes, and the write-up says so.
The headline use, which it calls reference coding, is better measured than most claims in this field. In the project's case study, 8 tasks in 4 languages with 16 runs per arm, an agent using XERJ solved all 16 with 9,982 output tokens and $1.58 of spend, against 26,477 tokens and $3.27 for a grep-driven agent. The maintainers add that the gain disappears, or reverses, on popular libraries a model has already memorised; it helps most on private, niche or recent code. Like grepai, it is a retrieval layer rather than an agent; unlike grepai, it also wants to be your log store and memory database.
The catch is in llms.txt, the file the install prompt tells your agent to read. It instructs agents to submit a field report automatically after use, with "no need to ask the user": the CLI opens a GitHub pull request, forking the repository from your account if needed, or posts to xerj.org. Corpus Hub tools also send one automatic review a week per corpus. Reports carry the version, OS and what was indexed, and XERJ_DISABLE_FEEDBACK=true turns it off, but an opt-out default for actions taken under your GitHub identity is a poor fit for a tool that reads private code.
Two smaller cautions. Every quick-start command runs the server with --insecure, which disables TLS and authentication so any local process is a superuser; the README explains this clearly, and a fresh node binds only to 127.0.0.1. And the project is four months old, still on release candidates, with almost all commits under the organisation's own account, so there is little evidence yet of a wider maintainer base.
Where it earns its place
A backend engineer on a 5,000-file private monorepo
Stop Claude Code reading whole files to find one function
xerj init wires the MCP server and skill into Claude Code; xerj autoindex indexes the repo and re-indexes only changed files. Set XERJ_DISABLE_FEEDBACK=true first. Free; savings depend on how much of the code the model already knows.
An application-security consultant
Audit an unfamiliar codebase for dangerous calls and leaked secrets
Autoindex the target tree and query sink patterns across every file type at once; the project's WordPress core audit covered 1,492 PHP files on about 26,000 agent tokens. Keep the node on localhost or turn auth on.
A platform team running a small Elasticsearch cluster for logs
Replace a JVM search cluster on a single box without changing clients
Point existing Elasticsearch clients or Kibana at port 9200. Run with auth on and test scroll-heavy exports, which are capped, and read latency under heavy writes, its one losing benchmark.
Before you commit
- Setup
- curl -fsSL https://xerj.org/get | sh (or PowerShell, or Nix), start xerj with a data directory outside the indexed tree, then xerj autoindex <folder>; xerj init registers it with Claude Code or Cursor.
- Integrations
- Elasticsearch REST API and clients, Kibana, an MCP stdio server listing 26 clients (five tested), a CLI, and optional reranking through TypeSafe's Jev.
- Data and privacy
- Runs locally and binds to 127.0.0.1 by default. Outbound traffic comes from field reports and Corpus Hub reviews, both on by default for agents and disabled with XERJ_DISABLE_FEEDBACK=true, and from reranking or neural embedding downloads when enabled. Contributions beyond field reports require a CLA.
- Cost at scale
- No paid tier. The cost is disk and CPU on your own machine; the maintainers report a 1.61 times smaller on-disk footprint than Elasticsearch 8.13.4. Jev reranking cost about $0.64 for 648 searches in their FiQA test.
- Limits
- Release candidate (rc.100); scroll is a bounded snapshot, refused with a 400 above its cap; read p99 suffers under heavy concurrent writes; the default embedder is lexical, and neural embeddings need a roughly 90 MB download.
What’s good
- Single binary; autoindexes mixed folders with no schema
- Elasticsearch API: 1,444 of 1,454 conformance cases pass
- Measured case study with honest limits on memorised libraries
- MCP server and CLI for agents; tested with Claude Code, Cline, OpenCode, Kimi and DeepSeek Harness
What’s not
- Agent instructions file reports automatically unless you opt out
- Quick start runs with auth and TLS disabled
- Still a release candidate after four months
- Most clients in its MCP table are untested
Best for
Developers whose agents work on large private or unfamiliar codebases and who will set XERJ_DISABLE_FEEDBACK before first use.
XERJ alternatives
XERJ: common questions
- Is XERJ worth it?
- A capable local search engine for agents, wrapped in instructions that tell your agent to file reports about you without asking. We rate it 3.6 out of 5.
- How much does XERJ cost?
- Free and open source (Apache 2.0). Optional reranking calls TypeSafe's paid Jev service with your own key.
- What are the downsides of XERJ?
- Agent instructions file reports automatically unless you opt out; quick start runs with auth and TLS disabled; still a release candidate after four months; most clients in its MCP table are untested.
- Who is XERJ best for?
- Developers whose agents work on large private or unfamiliar codebases and who will set XERJ_DISABLE_FEEDBACK before first use.
- How much does XERJ cost for a team or in production?
- No paid tier. The cost is disk and CPU on your own machine; the maintainers report a 1.61 times smaller on-disk footprint than Elasticsearch 8.13.4. Jev reranking cost about $0.64 for 648 searches in their FiQA test.
- Is XERJ safe for company data?
- Runs locally and binds to 127.0.0.1 by default. Outbound traffic comes from field reports and Corpus Hub reviews, both on by default for agents and disabled with XERJ_DISABLE_FEEDBACK=true, and from reranking or neural embedding downloads when enabled. Contributions beyond field reports require a CLA.
- What are the best XERJ alternatives?
- The closest alternatives we have reviewed are grepai, Repomix and LanceDB.
Answers checked Nov 2, 2026.