Skip to main content
Glama

Local Code RAG

Local semantic code search for Claude Code: index folders (Unreal projects, engine source, anything), embed them on-device with FastEmbed/ONNX, store them in LanceDB, and expose hybrid (vector + BM25) search over MCP. A Tauri v2 desktop app manages the indexes.

crates/rag-core   indexing, chunking, embeddings, LanceDB storage, config, file locks
crates/rag-mcp    headless MCP server (stdio, rmcp)
src-tauri + ui/   Tauri v2 desktop manager (React + TS + Tailwind v4)

Build

Requirements: Rust ≥ 1.94, Node 22, protoc (LanceDB dependency). Windows: MSVC build tools + WebView2 (preinstalled on Win10/11).

# MCP server (CUDA by default; falls back to CPU at runtime if CUDA/cuDNN are missing)
cargo build --release -p rag-mcp
cargo build --release -p rag-mcp --no-default-features                       # CPU only
cargo build --release -p rag-mcp --no-default-features --features directml   # AMD/Intel DX12 GPUs

# Desktop app: bundles rag-mcp + the ONNX Runtime CUDA provider DLLs next to the exe
npm ci
npx tauri dev                       # development (run `cargo build -p rag-mcp` once first)
npx tauri build                     # installer in target/release/bundle
$env:RAG_FEATURES="none"; npx tauri build      # CPU-only bundle

CUDA: the user's machine needs the CUDA 13 runtime + cuDNN 9 (for CUDA 13) on PATH (not bundled, ~1 GB) and a recent NVIDIA driver. The pinned ONNX Runtime (ort 2.0.0-rc.15) only ships CUDA 13 prebuilt binaries; with CUDA 12 installed it silently runs on CPU. The GUI's chunks/sec readout shows whether the GPU is actually used (CPU is ~10–30x slower).

ONNX Runtime binaries are downloaded at build time by ort; models are downloaded from Hugging Face on first use into <data dir>/models.

Related MCP server: MCP Codebase RAG Server

Data layout

%APPDATA%\local-code-rag\ (Windows) · ~/.local/share/local-code-rag/ (Linux) · override with LOCAL_CODE_RAG_HOME.

config.json              global settings + per-index config (schema below)
manifests/<id>.json      relative_path -> {blake3, size, mtime}
locks/<id>.lock          fs2 advisory lock (GUI indexing vs MCP search/auto-sync)
lancedb/idx_<id>.lance   chunks: path, start_line, end_line, text, vector
models/                  FastEmbed ONNX cache

config.json follows the blueprint schema (active_embed_model, execution_provider = auto|cuda|directml|cpu, indexes[] with root_path, embed_model, allowed_extensions, ignore_dirs, chunk_size, chunk_overlap, auto_update_on_search, auto_update_cooldown_hours, last_indexed_at, chunk_count), plus global embed_batch_size, reranker_model, rerank_candidates, auto_sync_max_files, snippet_max_lines and description / max_file_bytes (per index).

How indexing works

  1. Walk root_path with ignore (respects .gitignore; skips ignore_dirs by name at any depth, case-insensitive; filters by extension; skips files > max_file_bytes).

  2. Hash with blake3 — files whose size + mtime match the manifest are not re-read, so a no-change sync of a large tree is just a directory walk.

  3. Diff against the manifest → delete chunks of removed/modified files.

  4. Chunk new/modified files: tree-sitter AST splitting for C/C++/HLSL (.usf/.ush), Rust, C#, Python; plain-text splitter otherwise. Each chunk keeps its 1-based line range.

  5. Sort chunks by length (minimises ONNX padding), embed in batches (embed_batch_size, default 32), flush to LanceDB every 2048 rows. The embedded text is prefixed with the file path for context.

  6. Rebuild the full-text index on text (no stemming / stop words — code-friendly), a scalar index on path, and an ANN vector index once a table exceeds 100k rows; then compact.

Changing an index's model or chunk settings changes its fingerprint → next sync is a full rebuild (the GUI warns and offers to do it now). Cancelling keeps all completed files; the next sync resumes.

Search embeds the query (with the model's query prefix for BGE/Nomic), runs LanceDB hybrid search (vector + BM25, fused with Reciprocal Rank Fusion) over the top rerank_candidates (30), re-scores them with a cross-encoder (reranker_model, default jinaai/jina-reranker-v2-base-multilingual, null disables), and returns the top limit. Each result is trimmed to the snippet_max_lines (40) lines that best match the query terms (identifiers are split on :: and camelCase), labelled with the original chunk range.

Auto-sync on search only runs inline when at most auto_sync_max_files (200) files changed and no full build is needed. Otherwise the search answers from the existing data with a note asking for "Index now" in the app (or reindex_path), so a tool call never blocks on a multi-minute build.

Note: current LanceDB versions ship their own native BM25 inverted index instead of the older Tantivy-based one; the behaviour is the same (hybrid + RRF), with no extra dependency.

MCP tools

Tool

Purpose

list_indexes()

ids, names, roots, models, chunk counts, last update, status

search_index(index_id, query, limit=6)

hybrid search; incremental auto-sync first if auto_update_on_search and past cooldown

reindex_path(index_id, force_full=false)

manual sync

Logging goes to stderr only (RUST_LOG to adjust). If the GUI is indexing, auto-sync is skipped and search proceeds against the last committed data (LanceDB reads are MVCC-safe).

Register (or use the Install to Claude Code button, which also generates a CLAUDE.md snippet):

claude mcp add --transport stdio --scope user local-code-rag -- "C:\Program Files\Local Code RAG\rag-mcp.exe"
claude mcp list

Per-project AGENTS.md

Each index card has an AGENTS.md button that generates a paste-ready paragraph for that project's AGENTS.md / CLAUDE.md: its index id, root path and file types, when to call search_index vs. reindex_path, and the other indexes (e.g. engine source) the agent can also search. Generated by rag_core::agents_md::agents_md_snippet.

Tests

cargo test -p rag-core                        # unit tests (walker, manifest diff, chunker, model resolution)
cargo test -p rag-core -- --ignored           # end-to-end: downloads bge-small, indexes, searches, syncs

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      A local, SQLite-backed code index for Claude Code, exposed over MCP, enabling targeted code retrieval without external APIs.
      112 PyPI
      1
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables semantic code search over local repositories, providing tools like semantic_search and list_indexed_repos to Claude Code, so users can find relevant code sections via natural language instead of grep.
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables MCP clients such as Claude Desktop and Claude Code to query an indexed source repository through path-guarded tools for hybrid keyword-and-vector code search, file snippets, symbol explanations, reference lookups, and test-skeleton suggestions. Serves those results over stdio with cited file:line locations so agents can answer questions about a codebase and drive grounded code review.
      MIT