Sverklo (@sverklo) header
Sverklo (@sverklo) profile picture

Local-first code intel for AI agents. MCP server, symbol graph, blast-radius. 43× fewer tokens than grep on the public bench. MIT, zero-config.

Joined May 2026
Photo by Sverklo (@sverklo)
Sverklo gives coding agents local repo intelligence before they edit: symbols, callers, dependencies, and git-tied memory with public benchmark methodology. Local-first, MIT/open source. sverklo.com/benchmarks/
1
53
Sverklo is taking one opt-in maintainer question next week: raw evidence → answer → correction, if needed. No performance claims. No competitor dunking. Just source paths and failure modes. Reply with repo + pinned commit + question.
Sverklo’s pre-edit checklist: 1. Which symbol changes? 2. What calls it? 3. Which test constrains it? 4. Which commit is this? If an answer is fuzzy, I slow down. What is on yours? Try Sverklo:
Sverklo is built around one rule: when an agent is wrong, you should be able to inspect why it believed that. A visible evidence trail beats a confident paragraph. What is your minimum bar before you accept an agent edit?
Sverklo needs `grep-better` cases: public-repo questions where `rg` is faster, clearer, or more reliable than code intelligence. That is evidence, not a heckle. Best format: repo / pinned commit / question. We only name projects with maintainer approval.
I’m building Sverklo for the question maintainers actually get: “Where does this behavior come from?” If the answer is scattered across source, a test, and old discussion, send a public repo + pinned commit + question. Hard cases > flattering demos.
Sverklo is a local-first MCP server for coding agents. A list of “20 relevant files” is not an explanation. For a codebase question, I want: symbol → callers → tests → decision. What evidence do you need before approving an agent edit?
2
Sverklo is a local-first MCP server for coding agents. I tested it on one Vite HMR question. The first answer was wrong: sourcemap code ranked above the real HMR path. That is the correction we chose to publish.
1
1
We added a regression and changed path ranking. In a local warm-cache rerun of a pinned Vite commit, `updateModules` ranked #1 and the false lead fell to #97. Local evidence, not a benchmark or Vite endorsement.
1
Try the exact correction, or send a public repo / commit / question where grep wins. We will publish the outcome either way: #MCP
New rule for Sverklo: No “understands your codebase” claim without: • a pinned commit • a source path • a way to show the answer wrong We learned it from a Vite miss. The correction comes tomorrow. What would you add?
1
1
16
Java impact analysis is better now. Sverklo now counts Java type refs: Foo<T> implements Foo Foo.class Fix shipped. CI green. github.com/sverklo/sverklo
5
The coding-agent context problem is not just "fit more repo into the prompt." It is knowing which evidence to fetch before editing: - symbols - callers - tests - dependencies - stale memory Bigger context helps. Better retrieval changes the failure mode
4