Overview

zebrarag is a local-first semantic code index and MCP server. It parses your codebase with tree-sitter, chunks it by structure, embeds everything on your own hardware, and exposes fast query tools for AI coding agents — no cloud, no API keys.

How it works

your repo
  ↓ tree-sitter
AST symbols + edges
  ↓ chunking
semantic chunks (symbol-aware or recursive)
  ↓ Candle (BERT/Jina)
float32 embeddings  →  TurboQuant (int8)
  ↓ LanceDB + usearch
vector + full-text index
  ↓ MCP stdio server
AI agent tools: searchQuery · searchDep · fileTree · …

Architecture

zebrarag is a workspace of focused Rust crates:

Crate	Role
`zrag-daemon`	Background process, handles IPC requests
`zrag-pipeline`	Index + search pipeline orchestration
`zrag-dsl`	Tree-sitter code parsing, chunking, call graphs
`zrag-embed`	Candle-based embedding engine (BERT, Jina, BGE, …)
`zrag-ann`	ANN index (usearch) + TurboQuant reranker
`zrag-store`	LanceDB storage layer
`zrag-protocol`	IPC request/response types
`zrag-tree-sitter`	Language detection + tree-sitter frontends
`zrag-rerank`	TurboQuant reranking (CPU & GPU)
`apps/zebrarag`	CLI, TUI, and MCP server entry point

Key properties

Local-first — embeddings run on your CPU, Metal (macOS), or CUDA GPU
Incremental — only changed files are re-embedded; daemon stays alive
MCP-native — ships as a stdio server; works with Claude Code, Cursor, Zed, Windsurf, Continue.dev
Multi-language — Rust, TypeScript, JavaScript, Python, Go, Dart, Solidity, OCaml
12 embedding models — from 22M-param MiniLM to 567M BGE-M3, auto-downloaded

Quick start

cargo install --git https://github.com/hicaru/zebra_tree_indexer

# index the current project
zebrarag index .

# search from the terminal
zebrarag search "retry logic"

# start as MCP stdio server for AI agents
zebrarag --mcp

See install for full setup and MCP config snippets.