Single SQLite file per repo. No external databases, no Docker, no cloud services. Just a .srclight/ directory alongside your code.
Local embeddings via Ollama. Your code never leaves your machine. Works on air-gapped networks and private repos.
Keyword (FTS5 trigram + porter stemmer) fused with semantic embeddings via Reciprocal Rank Fusion. The best of both worlds.
ATTACH multiple SQLite databases and UNION across schemas. Query your entire codebase as one, with batch rotation for the 10-ATTACH limit.
Precise symbol extraction: functions, classes, methods, interfaces, structs. With call graph edges and type hierarchies.
Optional CuPy backend for GPU-resident vector cache. Sub-3ms semantic queries over tens of thousands of symbols.
One pip install. No compilation, no system deps beyond Python 3.10+.
pip install srclight
Point at any repo. Tree-sitter parses symbols, FTS5 indexes everything, optionally generate embeddings.
srclight index /path/to/repo
srclight index --embed \
qwen3-embedding /path/to/repo
Start the MCP server. Connect from Claude Code, Cursor, Windsurf, or any MCP client.
srclight serve \
--workspace myworkspace