sverklo, from Sverklo, is a Model Context Protocol server that gives AI coding agents persistent, structural "repo memory" for large codebases. The tool builds a local symbol graph and exposes 37 specialized tools for navigation, impact analysis, and code auditing to supply richer context and reduce token consumption. Key capabilities include hybrid semantic search, blast-radius analysis, diff-aware reviews, and Git-pinned decision memory. Software engineers and AI researchers who require private, agent-aware code context benefit most.

What tasks can you actually use the tool for?

The server supplies agent-facing code intelligence rather than plain text search: it creates a local symbol graph for relationship mapping and offers 37 specialized utilities that support navigation, impact analysis, and code auditing. Use cases include targeted code navigation, finding load-bearing symbols, performing refactoring impact studies, and automated detection of dead code and circular dependencies. These functions aim to convert raw repository files into structured context usable by AI coding agents.

How accurate are the outputs for code intelligence tasks?

Output fidelity rests on a hybrid retrieval stack combining BM25, vector embeddings, and PageRank, which the developer publishes benchmark results for across open-source repositories. The tool generates blast-radius metrics via sverklo_impact and assigns risk scores for pull-request diffs, and it flags circular dependencies and dead code during audits. Those features reduce token use by delivering targeted structural excerpts instead of large raw file dumps.

What inputs and environment does the server require?

Installation and runtime depend on Node.js version 24 or higher and an environment that supports MCP hosts. The parser layer uses tree-sitter for incremental parsing and supports more than a dozen languages, explicitly including TypeScript, JavaScript, and C#. Install options include a global npm package or running directly with npx, and the system accepts local repositories for indexing rather than relying on external APIs.

Does it fit into existing developer workflows and privacy needs?

The developer designed the server to operate locally with bundled embeddings and on-machine indexing, which avoids cloud uploads and API keys. Git-pinned decision memory lets agents retain repository-specific context between sessions for reproducible agent state. Because all indexing and embeddings run locally, teams must plan for initial indexing time and disk or CPU use, but the tool integrates into agent-driven code review workflows without external data sharing.

A practical choice for teams prioritizing private, structural agent context

The tool is a practical option for engineering teams integrating AI agents into code review and refactoring workflows, provided they accept an upfront environment setup and indexing period. Its outputs suit automated reasoning pipelines but should be paired with human review when decisions affect correctness or security. For teams that value private, agent-aware repository context and can allocate local compute, the tool delivers measurable workflow value.