Investigation: evaluate Understand-Anything for onboarding onto large codebases

[xmanjack]

Context

We've added Understand-Anything as a git submodule in our research repo and want to evaluate it as a tool for quickly understanding large/unfamiliar codebases via its knowledge-graph pipeline.

Goals of this investigation

• Pipeline & graph quality — Understand how the multi-agent pipeline analyzes a project and how accurately the knowledge graph captures files, functions, classes, and dependencies.
• Scale & cost — Evaluate behavior, runtime, and token/cost on a large repo (100k+ LOC), plus any incremental/re-analysis support.
• Integration & dashboard — Assess the Claude Code plugin install flow, cross-tool support (Codex, Cursor, Copilot, Gemini CLI), and the interactive dashboard's usefulness for exploration and Q&A.

Questions to answer

• How well does it handle our primary languages/frameworks?
• What data leaves the machine, and what are the privacy/security implications?
• How does graph accuracy degrade on very large or polyglot repos?

Deliverable

A short write-up summarizing findings and a go/no-go recommendation.

[Lum1104]

Sounds good, I am insterested in large scale evaluation too. Can you scale more, like to 1000k loc, or a monorepo? Feel free to contact me at lin.yuxiang.contact@gmail.com