xerj: Local-First Search for Code and AI Agents

Overview

XERJ is a local-first search engine and MCP server for AI agents, built in Rust. It indexes code, documents, logs and structured files, then serves BM25 full-text, vector and hybrid search through an Elasticsearch-compatible API. The goal is to give agents relevant passages instead of making them read whole files or rely on grep alone.

It is aimed at developers using Claude Code, Cursor or other MCP clients, as well as teams that want local search, RAG or agent memory without a separate search service. It ships as a single binary for Linux, macOS and Windows. The default lexical embedder works offline; neural embeddings are optional.

Key Features

• Automatically indexes folders containing source code, documents, logs, CSV, JSONL, PDFs and other supported formats.
• Parses source code with tree-sitter to provide symbols and line locations alongside full-text search.
• Provides BM25, vector and hybrid search, with query syntax compatible with Elasticsearch clients.
• Exposes search and memory tools to agents through an MCP server included in the binary.
• Supports persistent agent memory through namespace-based memory endpoints.
• Offers code definition lookup and reference-coding search for finding implementations in related repositories.
• Includes incremental indexing that skips parsing unchanged files.
• Provides an optional neural embedding mode; the default lexical embedding mode is offline.

Use Cases

• Reference coding: Developers can index related open-source projects and retrieve implementation examples before writing code, especially when working with unfamiliar or less common APIs.
• Codebase Q&A: Coding agents and engineers can search a local repository for relevant passages and definitions rather than opening files one by one.
• Local RAG and agent memory: Teams can index their own documents and let agents retrieve information across sessions without relying on a separate vector database.
• Log and document search: Operators can query mixed-format logs and files through one Elasticsearch-compatible interface.
• Security review: Auditors can search a code tree for risky calls, secrets and sink patterns across supported file types.

Getting Started

Install and index a project:

curl -fsSL https://xerj.org/get | sh
xerj --insecure --data-dir ~/xerj-data &
xerj autoindex ~/my-project

--insecure disables TLS and API-key authentication, so use it only for a local development node. See the README and installation guide for authenticated setup, MCP configuration and platform details.

Alternatives

• codebase-memory-mcp: It indexes repository structure into a code graph for architectural queries, while xerj offers BM25, vector, and hybrid search across files.
• Koshi: Koshi combines project memory and code retrieval for coding agents, while xerj centers on general-purpose file indexing and hybrid search.
• open-index: Open Index stores structured concepts and relationships for agent memory, while xerj focuses on full-text, vector, and hybrid search over files.

Considerations

• The project is less than six months old and changing quickly. Recent commit and release activity is high, but that pace also means behavior may change as it matures.
• Most code contribution is concentrated with one person, despite a wider contributor base. This is a project-continuity consideration, not a sign of inactivity.
• The Elasticsearch compatibility suite still has documented gaps. XERJ is compatible at the API level, but is not an Elasticsearch internals reimplementation; check the compatibility documentation for your use case.
• A fresh --insecure node binds to localhost, but that mode has no authentication or TLS. Every caller able to reach the port can access and delete indexes. Use the documented authenticated mode for shared hosts.
• Neural embedding is optional and requires model files to be available locally before semantic operations. The default lexical mode is offline.
• Optional reranking sends result text to an external service when configured. It is disabled unless enabled with a provider key; review the reranking documentation before using it.
• The project is licensed under Apache-2.0. Review the license when distributing or modifying it.