GitHub Repository

Deep code indexing MCP server for AI agents. 25 tools: hybrid FTS5 + embedding search, call graphs, git blame/hotspots, build system analysis. Multi-repo workspaces, GPU-accelerated semantic search, 10 languages via tree-sitter. Fully local, zero cloud dependencies.

46 starsPython

Srclight – Deep code indexing MCP server (FTS5 and Tree-sitter)

Name: Srclight – Deep code indexing MCP server (FTS5 and Tree-sitter)
Availability: InStock
Author: srclight

by srclight·Feb 24, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●●BangerWizardryBig BrainZero to One

Tree-sitter + FTS5 + MCP = tokens saved for AI agents to actually code, not search.

Strengths

•Hybrid RRF search (keyword + embedding fusion) with GPU acceleration—genuinely clever ranking strategy, not just chaining APIs.
•Offline-first with local Ollama embeddings; zero API calls means code never leaves your machine, real win for security-conscious teams.
•25 specialized MCP tools covering callers, relationships, git hotspots, and build awareness—breadth shows domain depth, not feature bloat.

Weaknesses

•Requires Python 3.11+, Git, optional Ollama setup—higher onboarding friction than plugins that ship with Cursor/Claude.
•Market timing: MCP ecosystem still nascent; adoption depends on Claude Code and Cursor integration maturity.

Post Description

Hi HN, I built srclight because AI coding agents (Claude Code, Cursor, etc.) waste 40-60% of their tokens just searching for code and understanding structure. grep and glob aren't enough when you need to know who calls a function, what changed recently, or find code by concept.

srclight indexes your codebase with tree-sitter and builds three FTS5 indexes (symbol names, source code via trigram, docstrings via Porter stemmer), a relationship graph (callers, callees, inheritance), and optional embedding vectors for semantic search. Everything lives in a single SQLite file per repo — no Docker, no Redis, no cloud APIs. Your code never leaves your machine.

Key features: - 25 MCP tools covering symbol search, relationship graphs, git change intelligence (blame, hotspots, WIP), build system awareness (CMake, .csproj), and semantic search - Hybrid search: RRF fusion of keyword + embedding results - Multi-repo workspaces: SQLite ATTACH+UNION across repos - GPU-accelerated vector cache (~3ms queries on 27K vectors via cupy) - 10 languages: Python, C, C++, C#, JS/TS, Dart, Swift, Kotlin, Java, Go - Incremental indexing with git hooks (post-commit, post-checkout) - Works with any MCP client (Claude Code, Cursor, Windsurf, Cline, VS Code)

I surveyed 50+ MCP code search servers across all the registries — most are grep wrappers or need cloud embedding APIs. srclight is the only one combining FTS5 + semantic search + git intelligence + multi-repo workspaces in a single pip install.

Install: pip install srclight GitHub: https://github.com/srclight/srclight