gitnexus

GitNexus (by Akon Labs) is a code intelligence engine that turns a repository into a queryable knowledge graph. Instead of handing an AI agent raw graph edges and hoping it explores enough, it precomputes structure at index time — clustering, call-chain tracing and confidence scoring — so a single tool call can return complete context about a symbol, its dependents and the execution flows it participates in. How it works Running `npx gitnexus analyze` from a repo root indexes the codebase, installs agent skills, registers hooks for supported editors and generates AGENTS.md / CLAUDE.md context files. `npx gitnexus setup` writes the MCP configuration so editors such as Claude Code, Cursor, Codex, Antigravity, Windsurf, OpenCode, CodeBuddy and Qoder can query the graph. Parsing uses Tree-sitter (native bindings for the CLI, WASM for the browser), and indexes are stored in LadybugDB — native and persistent for the CLI, in-memory WASM for the web UI. Two usage modes The CLI + MCP path is aimed at daily development: index locally, then let the agent query the graph. The Web UI (hosted at gitnexus.vercel.app or self-hosted) offers a visual graph explorer plus AI chat, limited by browser memory unless a backend is used. A bridge mode, `gitnexus serve`, lets the web UI detect a local server and browse CLI-indexed repos without re-uploading or re-indexing. MCP surface Seventeen MCP tools are exposed (15 per-repo plus 2 group-level), including `query` (hybrid BM25 + semantic + RRF search), `context` (360-degree symbol view), `impact` (blast-radius analysis), `trace` (shortest directed path between symbols), `detect_changes` (git-diff impact mapping), `rename`, `cypher` for raw graph queries, plus route/tool/shape/API-impact helpers and PDG-based `explain` and `pdg_query` for taint and statement-level dependence analysis. Companion MCP resources provide repo lists, clusters, processes, schemas and group contracts, and two prompts (`detect_impact`, `generate_map`) guide pre-commit analysis and architecture documentation. Agent skills and workflows Skills are installed into `.claude/skills/` and mirrored to `.agents/skills/` when present, covering exploring, debugging, impact analysis, refactoring, CLI usage, PDG queries and taint analysis. Slash-command style workflows include plan, work, review and a full plan→work→review pipeline. With `--skills`, Leiden community detection identifies functional areas and generates per-area project skills describing key files, entry points and cross-area connections. Deployment and operations One-click deployment blueprints exist for Render (a private server service plus a public web service with token-based auth) and RepoCloud. The README documents install caveats around npm 11, optional Tree-sitter grammars, proxy-restricted environments and embedding runtime downloads, as well as a read-only MCP mode and an allowed-repos policy for restricting which repositories agents may access. Licensing is PolyForm Noncommercial, with a commercial SaaS/self-hosted offering from the vendor.