Understand Anything codebase analysis tool
/Understand-Anything
Understand Anything is a plugin for Claude Code that turns a codebase into an explorable knowledge graph instead of a pile of files to read one by one. Splitting deterministic tree-sitter parsing from LLM-written summaries keeps the structural side reproducible, which is what actually makes the output trustworthy. Reach for it when joining an unfamiliar repo or sharing a graph with teammates; skip it if grep already answers what you need.
The Challenge of New Codebases
Understand Anything's README opens with a blunt scenario: you join a team, the codebase runs 200,000 lines of code, and there's no obvious place to start reading. Grep and a file tree only surface what you already know to search for, and comments or design docs, when a project has them, go stale the moment the code around them changes. That gap between code existing and someone being able to explain how it fits together is what a generated, explorable map is meant to close, instead of leaving a new contributor to trace imports by hand across dozens of files.
What Understand Anything Does
Understand Anything is a plugin for Claude Code that runs a multi-agent pipeline over your project and builds a knowledge graph covering its files, functions, classes, and dependencies. Tree-sitter extracts structural facts deterministically while an LLM layer adds plain-English summaries, architecture-layer tags, and guided tours, viewable in a dashboard that also works with Cursor, VS Code Copilot, Codex, and Gemini CLI.
Key Features for Codebase Insight
- ✓Interactive structural graph: each file, each function, and each class becomes a clickable node showing a plain-English summary, its relationships, and a guided tour.
- ✓Domain view: a separate mode maps the same codebase onto business logic — domains, workflows, and individual steps — rendered as a left-to-right graph, generated via /understand-domain.
- ✓Guided tours: auto-generated walkthroughs ordered by dependency, meant to teach the architecture in sequence rather than dumping the whole graph at once.
- ✓Fuzzy and semantic search across the graph, so a query like 'which parts handle auth?' surfaces relevant nodes without knowing exact file or function names.
- ✓Diff impact analysis via /understand-diff, showing which parts of the graph a pending change touches before you commit it.
- ✓Persona-adaptive dashboard UI that adjusts its detail level depending on whether you're a junior dev, a PM, or a power user, per the README.
- ✓Tree-sitter plus LLM hybrid pipeline: tree-sitter extracts imports, exports, function and class definitions, and call sites deterministically, while the LLM layer adds summaries, tags, and layer assignments.
Installing Understand Anything
For Claude Code itself, installation is two commands: /plugin marketplace add Egonex-AI/Understand-Anything, then /plugin install understand-anything. For every other supported platform — Codex, OpenCode, OpenClaw, Antigravity, Gemini CLI, Pi Agent, Vibe CLI, VS Code Copilot, Hermes, Cline, KIMI CLI, Trae, Nanobot, and Kiro — the README ships a one-line installer: curl -fsSL .../install.sh | bash on macOS/Linux, or install.ps1 via PowerShell on Windows. The script clones the repo to ~/.understand-anything/repo and symlinks the right files for whichever platform you pass, and it asks you to restart your CLI or IDE afterward. Cursor and VS Code with GitHub Copilot (v1.108+) auto-discover the plugin from .cursor-plugin/plugin.json or .copilot-plugin/plugin.json if you just clone the repo, no manual step needed. One thing worth knowing before you type a command: Codex uses a $ prefix instead of /, so it's $understand, not /understand.
Using the Analysis and Dashboard
Running /understand kicks off the multi-agent pipeline: it scans the project, extracts files, functions, classes, and dependencies, then writes the output into .ua/knowledge-graph.json; a project that already has an existing .understand-anything/ directory keeps using that path instead. /understand-dashboard then opens an interactive web dashboard, color-coded by architectural layer and searchable by node. From there, /understand-chat answers questions about the codebase in plain language, /understand-explain <path> goes deep on one particular file or function, /understand-diff checks what a pending change touches, /understand-onboard writes up a walkthrough for new joiners, and /understand-domain extracts the business-domain view. Re-running /understand is incremental by default, only re-analyzing changed files, and --auto-update wires that into a post-commit hook so the graph stays current on its own. On a large monorepo, scoping to a subdirectory (/understand src/frontend) keeps a single run smaller, and --language switches the generated summaries and dashboard UI to zh, zh-TW, ja, ko, or ru.
Who Should Use This Tool?
Understand Anything fits a developer who just inherited or joined a codebase they didn't write and wants a map before changing anything, especially on a team already living inside Claude Code, Cursor, Copilot, or one of the other supported platforms. It also suits teams that want to commit the graph once — the README points to GoogleCloudPlatform/microservices-demo as an example repo with a committed graph — so new hires or PR reviewers can open it with the standalone viewer and skip the pipeline entirely. Skip it if your project is small enough that a file tree and your editor's symbol search already answer most questions, or if nobody on the team uses any of the listed AI coding platforms to trigger the analysis in the first place.
Strengths
- ✓Works across a long list of AI coding platforms from one repo: native Claude Code plugin, plus a one-line installer covering a dozen-plus other CLIs and IDEs.
- ✓Hybrid pipeline keeps the structural side reproducible: tree-sitter facts (imports, calls, inheritance) don't change between runs, only the LLM-written summaries do.
- ✓Incremental re-analysis after the first run, only re-scanning changed files, keeps ongoing token costs down.
- ✓The graph is a plain committed JSON file, so a teammate can open it with the standalone npx viewer using only Node.js, no LLM calls or API key required.
- ✓MIT licensed, with no proprietary strings attached.
Considerations and Token Usage
- △The first full /understand run analyzes the entire codebase, and on a large project that first pass can burn through a lot of tokens; the README steers you toward a token plan/subscription or a local model for it.
- △Only the tree-sitter side of the pipeline is deterministic; the LLM layer writing summaries, tags, and architecture-layer assignments depends on whatever model you point it at, local or hosted.
- △Graphs above 10 MB need git-lfs tracking to commit cleanly, which is one more piece of repo tooling to adopt on a large project.
- △The README doesn't state a version number or give adoption details beyond stars and forks, so how this multi-agent pipeline holds up in day-to-day team use over time isn't something these facts confirm.
- △The standalone dashboard viewer needs a previously generated and committed .ua/ directory plus Node.js 18+; without both, there's nothing to open.
Alternative Approaches to Code Understanding
Common Questions
Understand Anything is a native plugin for Claude Code, and the README also lists one-line installers for Cursor, VS Code Copilot, Codex, OpenCode, Gemini CLI, Pi Agent, Vibe CLI, Hermes, Cline, KIMI CLI, Trae, Nanobot, and Kiro.
Yes. Running /understand-knowledge against a wiki built in the Karpathy LLM-wiki pattern produces a force-directed graph with community clustering, parsing wikilinks out of index.md before an LLM layer finds implicit relationships and entities.
The first full /understand run can burn through a lot of tokens on a large project, so the README suggests a token plan or a local model for that pass. After that, Understand Anything re-analyzes only changed files, and a monorepo can be scoped to a subdirectory to keep any single run smaller.
Yes. Committing the .ua/ directory (minus its intermediate/ and diff-overlay.json scratch files) lets teammates open the graph with a standalone npx viewer that needs only Node.js, no Claude Code or LLM calls required.
Understand Anything is released under the MIT License, as listed on its GitHub repository.
Yes. For privacy or enterprise setups, the README suggests pointing Understand Anything at a self-hosted model like Ollama instead of a hosted LLM.
Best use cases
- •Joining a team with a large, unfamiliar codebase and needing a starting map instead of reading files blind.
- •Committing the generated graph so new hires or PR reviewers can open it with the standalone viewer, skipping the pipeline entirely.
- •Checking what a risky change actually touches before committing it, via /understand-diff.
- •Turning a Karpathy-pattern LLM wiki into a browsable, clustered knowledge graph with /understand-knowledge.
- •Auto-generating a walkthrough document for people who just joined the team, via /understand-onboard.
Related repositories
Still deciding about Understand-Anything?
One click hands the question to an AI along with this page — see what it says about Understand-Anything.