sverklo
v0.20.0
/* comparison */

Sverklo vs Claude Context (Zilliz)

claude-context is Zilliz's MCP server: cloud-hosted vector search powered by Milvus, designed to give Claude an indexed view of large codebases. Sverklo is a local-first MCP that ships a real symbol graph alongside vector search — embedded SQLite, no external cluster, MIT-licensed. This page compares architecture, privacy, capability, and what each does best.

Feature comparison

Sverklo claude-context (Zilliz)
Storage Embedded SQLite (one file per repo) Zilliz Cloud or self-hosted Milvus cluster
License / Price MIT Free MIT code, but you pay Zilliz or operate Milvus
Where your code lives 100% local — never uploaded Embeddings + chunks live in Zilliz/Milvus
Operational footprint npm install -g sverklo — single binary Plus Milvus cluster or Zilliz account + API key
Retrieval Hybrid BM25 + vector + PageRank, fused with RRF Vector search (Milvus HNSW)
Symbol graph File-level + symbol-level dependency graph Not provided — chunks only
Blast-radius analysis sverklo_impact: callers, callees, test coverage Not available
PR review Risk-scored diffs with CI gate (--fail-on high) Not available
Memory of past decisions Bi-temporal, git-pinned, queryable at any past commit Not available
Audit / dead code sverklo audit — graded report, dead-symbol detection Not available
Tool surface 37 MCP tools (search, refs, deps, impact, memory, evidence, ctx_*, …) Index + search MCP tools
Scale ceiling Millions of chunks per repo on a laptop Billions of vectors on Milvus cluster
Open source Yes (MIT) Yes (MIT, but tied to Zilliz infra)

When to use each

Choose Sverklo when

  • You don't want to operate a Milvus cluster or pay Zilliz
  • Your codebase fits on a laptop (millions of chunks, not billions)
  • You need a symbol graph and blast-radius analysis, not just vector search
  • You want a queryable memory of past design decisions, git-pinned
  • You want diff-aware PR review with CI gates
  • Your code can't leave the machine — compliance, contract, or preference

Choose claude-context when

  • You're indexing many billions of vectors across a fleet of repos
  • You already operate Milvus or have a Zilliz Cloud account
  • You only need vector retrieval — no symbol graph, no diff review
  • You want a managed cluster instead of a per-machine SQLite database

Same problem, opposite architecture

Both projects want to give Claude (and other agents) a queryable view of the codebase. They diverge on where the data lives and how it's ranked. claude-context bets on a managed vector database: Milvus does the heavy lifting, the MCP layer is thin, and you scale horizontally by adding cluster capacity. Sverklo bets on local SQLite plus a symbol graph: the index is one file per repo, the ranker is hybrid BM25 + vector + PageRank, and the structural analysis (impact, refs, deps, risk-scored review) is built in.

For most individual developers and small teams, sverklo is the simpler answer — there's no cluster to operate and the symbol graph eliminates the entire class of "vector search returned the wrong file because the embedding got noisy" failures. For organisations already running Milvus at scale, claude-context is a clean MCP entry point.

Try Sverklo

MIT licensed. One command. Works with every MCP client.

click to copy