Overview

GitNexus (Akon Labs) is a zero-server code intelligence engine that indexes any codebase into a knowledge graph — every dependency, call chain, cluster, and execution flow — then exposes it through 17 MCP tools so AI agents never miss code. Its tagline is “the nervous system for agent context.” Unlike traditional Graph RAG that hands the LLM raw edges and hopes it explores enough, GitNexus precomputes structure at index time (clustering, tracing, scoring) so tools return complete context in one call.

Two deployment modes share the same graph engine: a CLI + MCP path (native Tree-sitter bindings, persistent LadybugDB storage, unlimited repo size) for daily development, and a Web UI (Tree-sitter WASM, in-browser LadybugDB, ~5k file limit) for quick exploration. A bridge mode (gitnexus serve) connects the two so the web UI can browse CLI-indexed repos without re-uploading.

Key Facts

  • Name: GitNexus (npm package gitnexus)
  • Vendor: Akon Labs (abhigyanpatwari)
  • License: PolyForm Noncommercial 1.0.0
  • Runtime: Node.js (native) or browser (WASM)
  • Storage: LadybugDB (formerly KuzuDB) — embedded graph database with vector support
  • Parsing: Tree-sitter ASTs (native bindings for CLI, WASM for browser)
  • Core innovation: “Precomputed Relational Intelligence” — structure computed at index time, not query time
  • MCP surface: 17 tools (15 per-repo + 2 group), 10 resources, 2 prompts
  • Agent skills: 12 standard skills auto-installed to .claude/skills/ and .agents/skills/, plus repo-specific skills generated via Leiden community detection (gitnexus analyze --skills)
  • Editor integrations: Claude Code (full: MCP + skills + PreToolUse/PostToolUse hooks), Cursor (full), Antigravity/Google (full), Codex (full), OpenCode (MCP + skills), CodeBuddy/Tencent (MCP + skills), Qoder/Alibaba (MCP + skills), Windsurf (MCP only)
  • Enterprise: SaaS and self-hosted via akonlabs.com
  • Web UI: https://gitnexus.vercel.app
  • ⚠️ No official cryptocurrency/token — any GitNexus token on Pump.fun or similar is unaffiliated

MCP Tools

ToolPurpose
queryProcess-grouped hybrid search (BM25 + semantic + RRF)
context360-degree symbol view — categorized refs, process participation
impactBlast radius analysis with depth grouping and confidence
traceShortest directed path between two symbols
detect_changesGit-diff impact — maps changed lines to affected processes
checkRead-only structural checks against the indexed graph
renameMulti-file coordinated rename with graph + text search
cypherRaw Cypher graph queries
route_mapAPI route map — components, endpoints, handlers
tool_mapMCP/RPC tool definitions and handlers
shape_checkValidate API response shapes against consumer property accesses
api_impactPre-change impact report for an API route handler
explainExplain persisted taint findings (source→sink flows)
pdg_queryStatement-level control/data dependence queries
list_reposDiscover all indexed repositories (paginated)
group_listList configured repository groups
group_syncRebuild a group’s Contract Registry and cross-repo links

Relationships

  • “uses” model-context-protocol — exposes 17 MCP tools, 10 resources, and 2 prompts for agent access
  • “operates-with” claude-code — deepest integration: MCP + agent skills + PreToolUse graph enrichment + PostToolUse stale-index detection hooks
  • “implements” knowledge-graph — builds a codebase knowledge graph as its primary artifact, with typed edges for calls, imports, heritage, and process participation
  • “implements” memory-retrieval-patterns — hybrid BM25 + semantic + RRF search pipeline in the query tool
  • “is-sibling-of” understand-anything — both are code-comprehension graph tools distributed as agent plugins; GitNexus emphasizes precomputed relational intelligence, Understand Anything emphasizes business-domain explanation
  • “is-sibling-of” graphify — both build graphs from folders with community detection; GitNexus uses Leiden clustering, Graphify uses EXTRACTED/INFERRED edge tags

Implications

GitNexus represents the precomputation turn in code intelligence for agents. Where traditional Graph RAG asks the LLM to explore raw edges at query time (4+ queries to answer “what depends on UserService?”), GitNexus computes clusters, traces, and confidence scores at index time so a single impact call returns the complete answer. The vendor frames this as “model democratization” — smaller LLMs work because the tools do the heavy lifting.

For the LLM Wiki ecosystem, GitNexus is notable for three reasons. First, it demonstrates that MCP is becoming the standard agent-tool interface: 8 editor integrations, all through the same MCP server. Second, its agent skills + hooks pattern (auto-installed to .claude/skills/, enriched by PreToolUse hooks) shows how a tool can shape agent behavior beyond just answering queries — it can inject context before the agent acts and detect staleness after. Third, its Leiden community detection for repo-specific skill generation mirrors the community detection this vault uses for graph analysis, but applies it to generate actionable agent instructions rather than navigation aids.

The PolyForm Noncommercial license is a meaningful constraint: it permits personal and research use but prohibits commercial deployment without an Akon Labs enterprise agreement. This positions GitNexus as an open-core product rather than a fully open-source tool.

Open Questions

  • Stars/forks not captured in snapshots (GitHub page rendered but stats not extracted).
  • All performance and integration claims are vendor-sourced; no independent benchmark captured.
  • How does the Leiden community detection compare in quality to Graphify’s approach for the same codebase?
  • Does the --pdg (program dependence graph) index add meaningful value over the standard call-graph for agent decision-making?
  • The Web UI’s ~5k file browser limit — is this a hard WASM constraint or a configurable threshold?

Sources

^[raw/external/github-com-gitnexus-7154a661.md]