GitNexus

GitNexus indexes a codebase into a knowledge graph and exposes it to AI coding agents through the Model Context Protocol (MCP), so an agent can see dependencies, call chains, clusters, and execution flows instead of inferring them from file reads. Two commands cover setup: gitnexus analyze indexes the repository, installs agent skills, registers hooks, and writes AGENTS.md and CLAUDE.md context files; gitnexus setup writes the MCP configuration for each detected editor.

Indexing runs as a multi-phase pipeline — file-tree structure, Tree-sitter parsing of functions, classes, methods, and interfaces, cross-file resolution of imports, calls, heritage, and receiver types, clustering into functional communities, execution-flow tracing from entry points, and hybrid search index construction. Because structure is precomputed at index time, a tool call such as impact or context returns a complete answer rather than raw edges the model has to explore.

Agents get 17 MCP tools, a set of read-only resources, two guided prompts, and skills installed into .claude/skills/ and .agents/skills/. The tools cover hybrid search, 360-degree symbol context, blast-radius impact analysis, shortest-path tracing between symbols, git-diff impact, multi-file rename, raw Cypher queries, API route and MCP tool maps, response-shape validation, and — on indexes built with --pdg — statement-level control and data dependence plus taint findings.

Indexes are stored in .gitnexus/ inside each repository and registered in a global ~/.gitnexus/registry.json, so one MCP server serves every indexed repo without per-project configuration. A browser build of the same engine runs Tree-sitter and LadybugDB as WebAssembly for in-browser exploration of repositories up to roughly 5,000 files, and gitnexus serve bridges the two so the web UI can browse locally indexed repos. Deployment options include Docker, a Render blueprint, and a Kubernetes controller.

Features

• Knowledge-graph index: functions, classes, imports, call chains, heritage, clusters, and execution flows extracted with Tree-sitter
• 17 MCP tools: search, context, impact, trace, detectchanges, check, rename, cypher, routemap, toolmap, shapecheck, apiimpact, explain, pdgquery, and group tools
• Precomputed structure: clustering, tracing, and scoring happen at index time so one tool call returns a complete answer
• 15 languages: TypeScript, JavaScript, Python, Java, Kotlin, C#, Go, Rust, PHP, Ruby, Swift, C, C++, Dart, and Zig, with per-language capability coverage
• Editor integration: gitnexus setup writes MCP config and skills, with hooks for the editors that support them
• Agent skills: exploring, debugging, impact analysis, refactoring, plan, work, and review skills, plus repo-specific skills generated per functional area
• Multi-repo registry: one MCP server serves every indexed repository, with lazy connections evicted after five minutes
• Program dependence graph: opt-in --pdg builds control-flow graphs for TypeScript and JavaScript to back taint and dependence queries
• Browser mode: a WebAssembly build indexes and visualises repositories client-side with Sigma.js and a LangChain ReAct agent
• Wiki generation: produces codebase documentation from the graph using an LLM API key, with configurable model, provider, and output language
• Hybrid search: BM25, semantic vectors, and reciprocal rank fusion over the indexed graph
• Local by default: CLI indexing makes no network calls and stores the index in a gitignored .gitnexus/ directory