Understand-Anything

Overview

Understand-Anything is an open-source Claude Code plugin that runs a multi-agent LLM pipeline over your codebase, building a structured knowledge graph and exposing it through an interactive web dashboard. Where most code-intelligence tools optimize for the agent’s query loop, Understand-Anything optimizes for the human who needs to understand the system — onboarding engineers, PMs writing system overviews, senior engineers planning refactors. The graph is a teaching artifact, not a query layer, and that distinction defines its place in the agent stack.

Key Capabilities

The five-agent pipeline runs project-scanner, file-analyzers (up to three concurrent), graph-builder, summarizer, and persona-tagger in sequence. The output is a .understand-anything/knowledge-graph.json file plus a local React 19 dashboard with React Flow + Dagre layout — every file, function, and class is a node, dependencies are edges, and nodes are color-coded by architectural layer. Each node carries a plain-English explanation the LLM wrote during indexing, which means semantic queries like “which parts handle auth?” return ranked results rather than fuzzy-matched filenames. Indexing is a one-time cost; the JSON output is committed to the repo so teammates skip the pipeline.

Use Cases

Three primary workflows: (1) onboarding new engineers to a legacy monorepo — the guided tour walks them through dependencies in order; (2) PMs and senior engineers writing system overviews without reading TypeScript line-by-line — the persona-adaptive UI surfaces architecture without code detail; (3) refactor planning — the diff-impact analysis shows what current changes affect before the commit lands. Where its closest comparable (colbymchenry/codegraph) optimizes for agent queries at the MCP layer, Understand-Anything optimizes for the operator-as-consumer.

Considerations

The LLM-built pipeline produces rich semantic context but is more expensive to index than a deterministic AST parser — operator reports put cost at $2–10 for medium repos, $20–50 for large monorepos, with 10–30 minute wall-clock per index pass. Re-indexing on every commit is impractical; the recommended pattern is to index once per major release branch and commit the JSON. The integration with non-Claude-Code agents (Codex, Cursor, Copilot, Gemini CLI) is via the committed JSON file rather than a tool-call interface, so the graph is read-only for those agents rather than queryable.

Who It’s For

Engineering teams onboarding new contributors to large codebases, PMs and stakeholders who need system overviews without reading code, and senior engineers planning risky refactors. Pairs naturally with codegraph (which queries the structure deterministically for agents) — install both for the full stack, with codegraph wired to the agents and Understand-Anything’s dashboard wired to the team.