Jonathon's AI Wiki

GitNexus — Zero-Server Code Intelligence Engine with Graph RAG

8 min read

Source: Abhigyanpatwari GitNexus README (2026-05-08) (github.com/abhigyanpatwari/GitNexus)

GitNexus is Abhigyan Patwari’s zero-server code-intelligence engine that turns any codebase into an interactive knowledge graph with a built-in Graph RAG agent. Drop in a GitHub repo or ZIP file via the browser-based Web UI (gitnexus.vercel.app), or run npx gitnexus analyze from a repo root for CLI + MCP integration with Claude Code, Cursor, and Codex. 16 MCP tools including hybrid search (BM25 + semantic), impact analysis, symbol context lookup, and multi-file rename. Tree-sitter AST parsing across 14+ languages (TypeScript, Python, Java, Go, Rust, etc.). Storage backed by LadybugDB; web runtime via WebAssembly. PolyForm Noncommercial license (free for personal/academic; commercial via akonlabs.com). 37,048 stars / 4,218 forks at fetch (2026-05-08); created August 2025, push as recent as 2026-05-08 — extremely active.

Key Takeaways

  • Zero-server is the load-bearing pitch. Most code-intel tools (Sourcegraph, Cody, GitHub Copilot Workspace) require either hosted backends or local server processes. GitNexus runs entirely in the browser via WebAssembly for the Web UI path. For CLI/MCP path, no server either — npx gitnexus analyze indexes locally and the editor connects via MCP. Privacy + portability + no infrastructure is the differentiator.
  • Two interface modes for two workflows.
    • CLI + MCP for Claude Code / Cursor / Codex integration: npx gitnexus analyze to index, gitnexus setup to configure MCP. The agent then has 16 MCP tools (hybrid search, impact analysis, symbol context, multi-file rename) for codebase-aware reasoning.
    • Web UI at gitnexus.vercel.app: drop a GitHub repo URL or ZIP file, get an interactive knowledge graph + a built-in Graph RAG agent for code exploration in the browser.
  • 16 MCP tools — what an LLM actually sees. Hybrid search (BM25 + semantic), impact analysis (“what breaks if I change this function?”), symbol context lookup, multi-file rename, plus the dependency-graph / call-chain / cluster / execution-flow tools the README enumerates. The MCP surface is the load-bearing integration point for Claude Code workflows — a Claude Code session with GitNexus MCP attached can answer “show me everything that calls processPayment” or “rename userId to customerId across the repo” without re-grepping each call.
  • 14+ language support via Tree-sitter. TypeScript, Python, Java, Go, Rust explicitly named. Tree-sitter AST parsing is the standard for cross-language code intelligence (used by GitHub, Neovim, Sourcegraph) — choosing it means GitNexus inherits high-quality parsers for the long tail of languages.
  • Six-phase indexing pipeline:
    1. File-structure mapping
    2. AST parsing via Tree-sitter
    3. Symbol resolution across files
    4. Functional clustering
    5. Execution-flow tracing
    6. Hybrid indexing combining BM25 and semantic search
  • LadybugDB graph storage. A graph database (presumably specialized for the workload) used for the knowledge-graph backend. Both Node.js native and WebAssembly-compiled — the same code runs in the browser or on a developer’s machine.
  • Both Claude plugin AND Cursor integration ship in-repo. First-class subdirectories gitnexus-claude-plugin/ and gitnexus-cursor-integration/ — the project is explicitly multi-tool, not single-editor. Composes with Claude Code plugins but not as a marketplace plugin (more likely a manual install).
  • Documentation discipline is unusual. 13+ markdown files at the repo root including AGENTS.md, CLAUDE.md, ARCHITECTURE.md (32KB), RUNBOOK.md, GUARDRAILS.md, DoD.md (Definition of Done), TESTING.md, MIGRATION.md, compound-engineering.local.md. Suggests commercial-grade discipline beyond a typical hobby project.
  • PolyForm Noncommercial license is a key constraint. Most tools cited in this wiki are MIT or Apache 2.0. PolyForm Noncommercial means: free for personal / academic / open-source projects, but commercial deployment requires a paid commercial license (via akonlabs.com — the company side). For WEO Marketly client work or any agency deployment, this requires a license; for personal exploration, it’s free. Worth flagging in any “use this” recommendation.
  • Author identity: Abhigyan Patwari is a real CS student / AI engineer with a credible profile (558 GitHub followers, 20 public repos, account from Feb 2023, GitHub bio matches the project’s interest area). akonlabs.com is the commercial-licensing entity. Not anonymous, not a stub.
  • The 37k-star number is striking for a ~9-month-old repo. Created Aug 2025, with active push activity through 2026-05-08 (one day before fetch). 4.2k forks suggests the indexing pipeline is genuinely useful, not just stargazer-curiosity. Worth pairing with a last30days research community-reaction check.

Where this fits in the wiki

  • Code-intelligence layer for AI-coding workflows. Sits between a developer’s editor and a codebase, providing the agent with structured queryable knowledge. Closest in shape to synthadoc (the Karpathy-pattern for documents) — same idea (build an indexed knowledge graph and let an agent query it), different domain (synthadoc indexes documents; GitNexus indexes code).
  • Composes with oh-my-claudecode (OMC) and Claude Managed Agents orchestration — gives those layers a structured codebase view for impact analysis and refactoring.
  • Adjacent to ai-website-cloner-template as another TypeScript-heavy multi-AI-tool project that ships its skill alongside a standalone Web UI.
  • Pairs with OpenSpec for spec-first workflows where the spec needs to anchor against the live codebase. OpenSpec defines what should exist; GitNexus shows what does exist.
  • Worth comparing against the six best Claude Code skills for business decision framework — does GitNexus pass the same vetting bar (publisher / last update / what it touches / dependencies / license / fit)?

Implementation

  • Tool/Service: GitNexus (abhigyanpatwari/GitNexus) — code-intelligence engine with Web UI + CLI + MCP server.
  • Setup:
    • Web UI: visit https://gitnexus.vercel.app, drop a GitHub repo URL or ZIP file. Browser-side WebAssembly indexes locally; no upload to any server.
    • CLI: npx gitnexus analyze from a repo root indexes the codebase locally and generates context files.
    • MCP for Claude Code / Cursor / Codex: gitnexus setup configures MCP automatically. Works with Claude Code via MCP server registration.
    • Docker: docker-compose.yaml + Dockerfile.cli and Dockerfile.web ship in-repo for containerized deployment.
  • Cost:
    • Free for non-commercial use under PolyForm Noncommercial.
    • Commercial licensing required for agency / company deployment — contact akonlabs.com.
  • Integration notes:
    • 14+ languages supported. Tree-sitter parsers for the long tail.
    • 16 MCP tools — once attached, Claude Code can call hybrid-search / impact-analysis / symbol-context / multi-file-rename without re-grepping.
    • LadybugDB stores the graph; same engine in Node.js native and browser WASM.
    • Cosign-signed Docker images + Kubernetes deployment policies documented (supply-chain hygiene).
    • The gitnexus-claude-plugin/ subdirectory is the recommended Claude integration path; gitnexus-cursor-integration/ for Cursor users.

Open Questions

  • PolyForm Noncommercial vs MIT in production. Most teams default to MIT-only deps. GitNexus’s PolyForm-NC license might prevent commercial deployment without a paid agreement. What’s the akonlabs.com pricing? Tier structure? Per-seat / per-repo / unlimited?
  • WebAssembly performance ceiling. WASM-based browser indexing for very large repos (100k+ files) — does it scale, or does it hit memory limits before the index completes? Not addressed in the README.
  • MCP tool stability. 16 MCP tools is a substantial surface. Are they stable enough to bake into a long-running production agent’s MCP config, or still experimental?
  • Tree-sitter language coverage in production. “14+ languages” — which 14? Are language-specific edge cases (TypeScript-with-JSX, Python-with-Cython, Rust-with-macros) handled cleanly?
  • Graph storage portability. LadybugDB is the storage layer. Is it queryable outside GitNexus, or locked into the project’s internal abstraction?
  • Comparison against Sourcegraph Cody / GitHub Copilot Workspace / Codeium. Those are hosted competitors. GitNexus’s “zero-server” framing is distinct, but for teams already on hosted code-intel, what’s the migration story / feature parity?
  • Browser-based privacy claim. “Runs entirely in your browser” implies code never leaves the developer’s machine. Worth verifying in the source — does any telemetry or analytics phone home? Not stated.

Try It

  1. Web UI smoke test. Visit https://gitnexus.vercel.app, drop in a GitHub repo URL of a project you know well. Look at the generated knowledge graph. Try a Graph-RAG-agent question like “show me everything that calls processPayment.”
  2. CLI quick start. cd <project> then npx gitnexus analyze — index the codebase locally, see what context files get generated.
  3. Wire MCP into Claude Code. gitnexus setup configures the Claude Code MCP server. Then in a Claude Code session, ask a code-impact question that requires cross-file traversal.
  4. Compare against grep + Read for the same question. Pick a refactor task you’ve done before. Run it once with GitNexus’s MCP tools available, once without. Note where the difference shows up.
  5. For commercial deployments, contact akonlabs.com about commercial licensing before depending on it in production. PolyForm-NC is not a commercial-default license.
  6. For curious tinkerers, the 13+ root-level markdown files (AGENTS.md / CLAUDE.md / ARCHITECTURE.md / GUARDRAILS.md / DoD.md / RUNBOOK.md / TESTING.md / MIGRATION.md / compound-engineering.local.md) are themselves a worked example of what good Claude-Code-targeted documentation looks like. Worth reading just for the convention bank.
  7. For multi-repo teams, the enterprise tier offers multi-repository support — useful for monorepos and microservice fleets where the cross-repo knowledge graph is the actual value.

Graph View

Backlinks