GitNexus indexes a codebase into a knowledge graph of its dependencies, call chains, functional clusters, and execution flows, then serves that graph to coding agents over the Model Context Protocol (MCP). The aim is that an agent editing a function knows what depends on it before a change ships, instead of discovering callers through a chain of searches.
Indexing runs in six phases: it walks the file tree, parses functions, classes, methods, and interfaces with Tree-sitter, resolves imports, calls, inheritance, and receiver types across files, groups symbols into functional communities, traces execution flows from entry points, and builds hybrid search indexes (BM25 plus semantic vectors fused with reciprocal rank fusion). Clustering, tracing, and confidence scoring happen at index time, so an impact query returns callers, clusters, and confidence in one tool call rather than handing the model raw graph edges to explore.
It ships in two forms. The CLI indexes repositories locally, stores each index in LadybugDB, and exposes a stdio MCP server that serves every indexed repository from one global registry, so a single setup covers all projects. The Web UI runs the same pipeline in the browser on WebAssembly builds of Tree-sitter and LadybugDB, offers a graph explorer and a chat agent built with LangChain, and is limited by browser memory (around 5,000 files) unless it connects to a local backend. A bridge mode lets the Web UI browse CLI-indexed repositories without re-indexing.
Setup auto-detects installed editors and writes their MCP configuration. Claude Code, Cursor, Antigravity, Codex, and Factory get MCP tools, agent skills, and hooks that enrich searches with graph context and flag a stale index after commits. OpenCode, CodeBuddy, and Qoder get MCP plus skills, and Windsurf gets MCP only. The indexer also writes AGENTS.md and CLAUDE.md context files and can generate one project skill per functional area of the codebase.
Parsing covers TypeScript, JavaScript, Python, Java, Kotlin, C#, Go, Rust, PHP, Ruby, Swift, C, C++, Objective-C, Dart, and Zig, with per-language differences in import, export, heritage, type-annotation, and entry-point resolution. An opt-in control-flow and program dependence graph (PDG) layer, currently for TypeScript and JavaScript, feeds statement-level dependence queries and taint analysis.
A wiki generator groups files into modules with an LLM and writes per-module pages and an overview from the graph, using a configurable model and provider (MiniMax by default, with OpenAI, Anthropic, Grok, or a custom provider base URL as alternatives). Local embeddings run through transformers.js and are opt-in. A commercial offering, hosted or self-hosted, adds pull-request review, auto-reindexing, a unified multi-repo graph, and OCaml support.
Features
- MCP tool set: 19 tools, including hybrid query, a 360-degree context view, impact blast radius with confidence, trace between two symbols, detect_changes on a git diff, multi-file rename, raw cypher, an API route_map, and shape_check for response shapes
- Resources and prompts: MCP resources list indexed repositories, clusters, execution flows, and the graph schema, and two prompts guide pre-commit impact analysis and architecture documentation with mermaid diagrams
- Agent skills: exploring, debugging, impact analysis, refactoring, PDG query, and taint analysis skills, plus plan, work, review, and full-pipeline slash commands, installed for Claude Code and for agents that read repo-local skills such as Codex
- Repo-specific skills: Leiden community detection splits the codebase into functional areas and generates one project skill per area, regenerated on each run
- Editor hooks: pre- and post-tool hooks in Claude Code and Codex enrich a search before it runs and prompt a reindex when the index goes stale after commits, with post-tool hooks for Cursor, Antigravity, and Factory
- Multi-repo registry: one MCP server serves every indexed repository, repository groups build a Contract Registry with cross-repo links, and linked worktrees share one index store
- Hybrid search: BM25 plus semantic vectors fused with reciprocal rank fusion, with opt-in local embeddings and a tunable vector distance cutoff
- Watch mode: re-indexes on local file changes with debounced incremental refreshes, and a watcher polls remote repositories on GitHub, GitLab, and Gitee
- Wiki generation: per-module documentation pages and an overview written by an LLM from the graph structure, in a chosen output language
- Web UI: an in-browser graph explorer and chat over a WebAssembly build of the pipeline, with no code leaving the machine
- Deployment targets: Docker images for the server and the Web UI, plus Render and RepoCloud deployments
- Control-flow layer: opt-in per-function control-flow graphs for TypeScript and JavaScript that feed statement-level dependence queries and taint findings
