Sverklo: A Local Code Intelligence Engine Built for AI Programming Assistants

Sverklo is an open-source local code intelligence tool designed for AI programming assistants. It addresses key limitations of existing tools by enabling deep understanding of codebase structures through AST parsing, PageRank ranking, and semantic search—all without cloud dependency. It supports MCP protocol integration with AI clients like Claude Code and Cursor, focusing on local-first, zero-config operation to protect code privacy while enhancing AI's code retrieval and analysis efficiency.

Existing AI programming assistants (e.g., Claude Code, Cursor) struggle with large codebases—wasting tokens on irrelevant files or missing key context. Cloud solutions (Augment, Greptile) raise privacy concerns for sensitive projects, while local tools (CocoIndex, Aider) lack features like graph ranking or MCP support. Sverklo fills this gap by combining local processing with advanced code analysis capabilities.

Sverklo uses four key technologies:

1. AST Parsing: Supports 10+ languages (TypeScript, Python, Go, etc.), extracts structured info (functions, classes, interfaces) for precise code understanding.
2. PageRank: Analyzes file dependencies to rank core modules higher, aligning results with developer intuition.
3. Local Semantic Embedding: Uses all-MiniLM-L6-v2 (ONNX runtime) for 384D vectors, no cloud needed.
4. Hybrid Search: Combines BM25 (keyword), vector similarity (semantic), and Reciprocal Rank Fusion with PageRank for optimal results.

Sverklo provides six MCP tools:

• search: Hybrid semantic search with token budget, path range, language, and symbol type filters.
• overview: High-level codebase view (PageRank-sorted core files and symbols).
• lookup: AST-based exact symbol (function/class/interface) lookup.
• find_references: Cross-file reference tracking for refactoring.
• dependencies: Import graph analysis (module coupling, traversal depth support).
• index_status: Index health check (file count, code block count, language support, progress).

Performance: For a 30-TypeScript-file project: index build (681ms including ONNX embedding), single search (<50ms), memory usage (~200MB). Incremental updates ensure real-time sync. Privacy: All processing is local—code never leaves the machine. Business: MIT-licensed open source (free core). Pro version adds session memory, Git integration; Team version for enterprise (shared memory, local deployment, management backend).

Value: Improves AI context quality (reduces hallucinations), lowers large project barriers (automates context selection), meets privacy needs (offline use for sensitive industries like finance/healthcare). Outlook: Sverklo represents a direction in AI programming tools—focusing on making AI 'understand' code better via traditional analysis + modern AI. It will become an infrastructure layer, enhancing AI assistants' potential without replacing them.