GitNexus is a self-hosted knowledge graph platform that indexes codebases into a deterministic graph of dependencies, call chains, and execution flows, enabling AI agents to navigate and edit code with greater accuracy and efficiency. The platform uses compiler-grade static analysis and graph-based clustering to provide agents with resolved execution flows rather than raw file searches, improving task solve rates while reducing token usage and cost.
Funding
Funding not disclosed
Founders
Product
Problem
AI coding agents that navigate codebases by reading files individually miss critical dependencies, break call chains, and make blind edits in large, interconnected systems. Traditional grep-based search is insufficient for understanding the full system architecture, leading to costly errors and inefficiencies in automated code modification.
Solution
GitNexus provides a self-hosted knowledge graph engine that indexes any codebase into a structured graph of every dependency, call chain, cluster, and execution flow. The platform uses Tree-sitter for deterministic parsing and resolution of imports, field types, and return types across repositories, eliminating guesswork from AI navigation. By precomputing the graph once and reusing it, GitNexus enables agents to walk resolved execution flows instead of grepping through files, resulting in a 68.4% solve rate on the DeepSWE benchmark compared to 37.0% for bare models. The system includes Leiden community detection for automatic architecture discovery, blast-radius analysis for change impact assessment, and multi-repo unified graphs for cross-service dependency tracking.
Target Audience
Primary customers are software engineering teams and enterprises using AI coding agents, particularly those working on large, multi-repo codebases where understanding system-wide dependencies is critical for safe automated code changes.
Features
- Compiler-grade static analysis with Tree-sitter that deterministically resolves imports, call chains, field types, and return types across the entire codebase
- Precomputed knowledge graph stored locally in a .gitnexus/ folder, enabling instant query responses without re-exploring code
- Leiden community detection algorithm that groups symbols into functional clusters with cohesion and modularity scores, revealing system architecture automatically
- Blast-radius analysis that maps git diffs to affected execution flows, tracing downstream callers by depth with confidence scores
- Multi-repo unified graph support that connects cross-repo edges, so breaking changes in one service surface in all dependent repositories
- Hybrid search combining BM25, semantic, and RRF ranking, plus 13 language support via Tree-sitter and MCP tools for Claude Code, Cursor, Codex, and Windsurf integrations
- 100% local and private by design, with indexing, storage, and queries running entirely on the user's machine or inside their network