Understand Anything

Why it's ranked here

The split between scripts and model is the reason to take it seriously. File enumeration, import resolution and structural extraction run as bundled scripts the agents are told not to reimplement, and a separate reviewer agent rejects any graph with dangling references or duplicate node IDs. That puts a checkable skeleton under the model's prose. It is MIT licensed, installs natively in Claude Code and through a one-line script for Codex, Gemini CLI, OpenCode and many other agents, and the maintainers ship a deterministic benchmark harness rather than speed claims.

What's good

Privacy handling is concrete: before the graph is written, absolute file paths are rewritten relative to the project, and any path outside it is cut down to its bare filename, so a shared graph does not leak a home directory or company folder layout. Reruns are incremental by default, re-analysing only changed files, and an optional post-commit hook keeps the graph current. The benchmark guide is candid about its limits: it measures only the deterministic stages, runs no model, counts no tokens, and says outright not to present it as an end-to-end result.

Tradeoffs

The first full analysis pushes the whole codebase through a model, and the README itself warns this can consume a significant number of tokens on large projects, recommending a subscription plan or a local model. Summaries, layers and tours are model output, so their quality follows whichever model you point it at; the reviewer agent checks graph structure, not whether a summary is right. Parser-based extraction covers ten languages; Swift, Kotlin, shell and PowerShell files fall back to the agent reading source itself. The headline installer pipes a remote script straight into the shell.

How to use it well

Best fit is an engineer landing in an unfamiliar, sizeable codebase who wants an overview before reading files one by one, or a lead preparing onboarding material for new hires. Run the first analysis on a subscription or local model, and scope it to a subdirectory in a large monorepo to keep cost down. Add an ignore file for generated and vendored folders. Treat the summaries as a starting map to verify against the code, not as documentation of record. It does not replace code search or review tooling for day-to-day changes.

Observed

License

MIT (LICENSE)

Language

TypeScript, organised as a pnpm workspace

Interface

Agent plugin driven by slash commands; Claude Code native, other agents via install script

Parsing

Tree-sitter grammars for structural extraction, declared in package.json

Storage

Local JSON files in a data directory inside the analysed project

Test runner

Vitest, declared in package.json

Benchmark scope

Deterministic stages only; no model calls or token counts (docs/benchmarks/large-monorepo.md)