smgrep

Natural-language search that works like grep. Fast, local, GPU-accelerated, and built for coding agents.

• Semantic: Finds concepts ("where do transactions get created?"), not just strings.
• GPU-Accelerated: CUDA on NVIDIA GPUs, Metal on Apple Silicon.
• Local & Private: 100% local embeddings. No API keys required.
• Auto-Isolated: Each repository gets its own index automatically.
• On-Demand Grammars: Tree-sitter WASM grammars download automatically as needed.
• Agent-Ready: Native MCP server and Claude Code integration.

Quick Start

1. Install

   cargo install smgrep
   

   Or build from source:

   git clone https://github.com/can1357/smgrep
   cd smgrep
   cargo build --release
   

   For CPU-only builds (no CUDA):

   cargo build --release --no-default-features
   

2. Setup (Recommended)

   smgrep setup
   

   Downloads embedding models (~500MB) and tree-sitter grammars upfront. If you skip this, models download automatically on first use.

3. Search

   cd my-repo
   smgrep "where do we handle authentication?"
   

   Your first search will automatically index the repository. Each repository is automatically isolated with its own index. Switching between repos "just works".

Coding Agent Integration

Claude Code

1. Run smgrep claude-install
2. Open Claude Code (claude) and ask questions about your codebase.
3. The plugin auto-starts the smgrep serve daemon and provides semantic search.

MCP Server

smgrep includes a built-in MCP (Model Context Protocol) server:

smgrep mcp

This exposes a sem_search tool that agents can use for semantic code search. The server auto-starts the background daemon if needed.

Commands

smgrep [query]

The default command. Searches the current directory using semantic meaning.

smgrep "how is the database connection pooled?"

Options:

Flag	Description	Default
-m <n>	Max total results to return	10
--per-file <n>	Max matches per file	1
-c, --content	Show full chunk content	false
--compact	Show file paths only	false
--scores	Show relevance scores	false
-s, --sync	Force re-index before search	false
--dry-run	Show what would be indexed	false
--json	JSON output format	false
--no-rerank	Skip ColBERT reranking	false
--plain	Disable ANSI colors	false

Examples:

# General concept search
smgrep "API rate limiting logic"

# Deep dive (more matches per file)
smgrep "error handling" --per-file 5

# Just the file paths
smgrep "user validation" --compact

# JSON for scripting
smgrep "config parsing" --json

smgrep index

Manually indexes the repository.

smgrep index              # Index current dir
smgrep index --dry-run    # See what would be indexed
smgrep index --reset      # Delete and re-index from scratch

smgrep serve

Runs a background daemon with file watching for instant searches.

• Keeps LanceDB and embedding models resident for fast responses
• Watches the repo and incrementally re-indexes on change
• Communicates via Unix socket (or TCP on Windows)

smgrep serve              # Start daemon for current repo
smgrep serve --path /repo # Start for specific path

smgrep stop / smgrep stop-all

Stop running daemons.

smgrep stop               # Stop daemon for current repo
smgrep stop-all           # Stop all smgrep daemons

smgrep clean

Remove index data and metadata for a store.

smgrep clean              # Clean current directory's store
smgrep clean my-store     # Clean specific store by ID
smgrep clean --all        # Clean all stores

smgrep status

Show status of running daemons.

smgrep list

Lists all indexed repositories and their metadata.

smgrep doctor

Checks installation health, model availability, and grammar status.

smgrep doctor

GPU Acceleration

smgrep uses candle for ML inference with optional CUDA support.

With CUDA (default):

Requires CUDA toolkit installed with environment configured: