Giving AI agents a memory palace for your codebase
GitNexus builds a knowledge graph of any repo so Cursor and Claude stop hallucinating your architecture.

What it does GitNexus indexes a codebase into a knowledge graph — dependencies, call chains, execution flows — then feeds that structure to AI agents via MCP. There’s a browser-based visual explorer for quick dives, and a CLI that wires the graph directly into Claude Code, Cursor, Codex, and friends so they actually know your code’s shape before they edit it.
The interesting bit The hook system is the quiet star. Claude Code gets PreToolUse hooks that inject graph context into searches, plus PostToolUse hooks that nudge the agent to reindex after commits. It’s not just “here’s some docs” — it’s a live, queryable map that updates as the code changes. The web UI runs entirely in-browser with a WASM build of their LadybugDB; the CLI uses native bindings for speed.
Key highlights
- MCP server with editor-specific integrations (Claude Code gets deepest hooks; Cursor, Antigravity, Codex, Windsurf, OpenCode supported)
- Browser UI at gitnexus.vercel.app — no install, no server, limited to ~5k files or backend mode
- Bridge mode (
gitnexus serve) lets the web UI browse CLI-indexed repos without re-uploading - One-command setup:
npx gitnexus analyzeindexes, installs skills, registers hooks, writesAGENTS.md/CLAUDE.md - PolyForm Noncommercial license; enterprise SaaS/self-hosted available separately
Caveats
- Install can be brittle: npm 11.x crashes on
npxinstall (known npm/arborist bug), and native grammar builds needpython3/make/g++unless you skip optional parsers - Web UI is memory-constrained; serious work wants the CLI
- License is noncommercial by default — commercial use requires separate enterprise deal
Verdict Worth trying if you’re tired of AI agents breaking call chains or missing dependencies in large repos. Skip it if you want a simple “explain this function” tool — DeepWiki already does that. The value is in the relationships, not the descriptions.