Both are code intelligence MCP servers. Both parse your codebase locally and expose tools to AI coding agents. They take different approaches to licensing, architecture, and feature scope. This page is an honest comparison to help you pick the right tool for your project.
Need the four-way overview? See the multi-product comparison covering sverklo vs Serena, GitNexus, and Greptile across six dimensions.
| Sverklo | GitNexus | |
|---|---|---|
| License | MIT Free for all use | PolyForm Noncommercial |
| Commercial use | Free, no restrictions | Requires enterprise license |
| Memory system | Bi-temporal, git-aware (memories tied to SHA) | None |
| PR review | Risk-scored with CI gate (--fail-on) | Change detection only |
| GitHub Action | Yes (open-source) | Enterprise only |
| Live file watching | Yes (incremental reindex) | Manual re-analysis |
| Install | npm install (pure JS, no native bindings) | npm install (native bindings required) |
| Languages | 10 via tree-sitter | 13 |
| Graph DB | SQLite + PageRank | KuzuDB + Cypher |
| Embeddings | Pluggable (ONNX default, Ollama for local models) | HuggingFace |
Both sverklo and GitNexus are on the public 5-baseline retrieval benchmark — 180 hand-verified tasks across 6 OSS codebases (express, lodash, sverklo, requests, flask, fastapi). Full ranking page. On the 2026-05-13 run sverklo wins overall F1 (0.58 vs 0.30); GitNexus wins lowest token cost (630 vs sverklo's 652, near-tie). Note: GitNexus maintainer @abhigyanpatwari has flagged the bench frames GitNexus as a "standalone retriever" when its design assumption is "graph-augmentation alongside the host agent's grep" — a sibling bracket measuring agent-loop cost is in discussion on sverklo-bench#2.
| baseline | F1 | P1 def | P2 refs | P4 deps | tokens |
|---|---|---|---|---|---|
| sverklo | 0.58 | 0.63 | 0.27 | 0.84 | 652 |
| jcodemunch-mcp | 0.29 | 0.52 | 0.01 | 0.33 | 1,907 |
| gitnexus | 0.30 | 0.35 | 0.00 | 0.27 | 630 |
GitNexus's strength is token cost — 630 average input tokens per task vs sverklo's 652 on the May 13 run, effectively a near-tie. The tradeoff is overall recall: GitNexus's F1 0.30 trails sverklo's 0.58 because the LLM-derived knowledge graph drops detail relative to a hybrid BM25 + embedding + PageRank stack. If your agent is operating under a tight context budget and you can tolerate the recall drop on def-lookup and deps, GitNexus's compression is real. If retrieval correctness is the priority, sverklo's overall lead translates directly.
Sverklo uses SQLite with PageRank-based scoring and a hybrid retrieval pipeline (BM25 + ONNX embeddings + graph centrality). Hybrid means lexical and structural signal both rank — neither tool plays the embedding lottery. This keeps the install simple and dependency-free. The trade-off: you don't get Cypher query syntax or native graph traversals.
GitNexus uses KuzuDB, a native graph database, which enables Cypher queries and structural graph visualization. The trade-off: native bindings can complicate installation on some platforms, and the PolyForm Noncommercial license restricts commercial use without an enterprise agreement.
MIT licensed. One command. Works with every MCP client.
click to copy