Code Knowledge Graphs as Agent Seatbelts

·https://github.com/abhigyanpatwari/GitNexus
repoaiclimcpcode-analysisknowledge-graphsagent-toolsdeveloper-toolsjoelclaw

GitNexus's MCP knowledge graph maps to joelclaw's agent-loop reviewer/debugging step: blast-radius checks before edits instead of hoping the model notices dependencies.

GitNexus indexes a codebase into a knowledge graph of dependencies, call chains, clusters, and execution flows, then exposes that graph to agents through a CLI, MCP, and a browser-based Web UI.

The clever bit is the positioning: not “chat with a repo,” but give the agent a nervous system for code context. DeepWiki explains code. GitNexus tries to make relationships queryable so agents can see what breaks before they touch it.

That matters for Cursor, Claude Code, Codex, and other coding agents because blind edits are where the wheels come off. A graph-backed agent can ask “what calls this?”, “what depends on this?”, and “what’s the blast radius?” instead of freeballing across a pile of files like a raccoon with commit access.

Useful for joelclaw because the system already cares about explicit reviewer/judge gates, dependency awareness, and not shipping busted call chains. The interesting move would be wiring graph queries into agent-loop review, PR checks, and debugging runs so “look around first” becomes enforced behavior, not vibes.

Key Ideas

  • GitNexus treats a repository as a knowledge graph, not just a search index or generated wiki.
  • The recommended path is CLI plus MCP, which gives coding agents local repo context without needing to ship code to a hosted service.
  • The Web UI is for quick exploration, demos, and repo chat, while local indexing is the serious daily-driver path.
  • Bridge mode via gitnexus serve lets the browser UI browse CLI-indexed repos without re-uploading or re-indexing.
  • LadybugDB handles persistent local storage for the CLI path and in-memory WASM storage for browser sessions.
  • Tree-sitter parsing is the foundation for language-aware graph construction.
  • The enterprise angle from Akon Labs points at automated PR blast-radius analysis, auto-updating code wiki, auto-reindexing, and multi-repo graphs.
  • The repo ships explicit agent-facing docs like ARCHITECTURE.md, RUNBOOK.md, and GUARDRAILS.md, which is the right smell for tools meant to be operated by agents.

Links

← All discoveries