Understand Anything：把任何代码库变成可交互的知识图谱
Understand Anything 是一个开源工具，用于将任意代码库生成可视化的交互式知识图谱，使开发者可以浏览、搜索和提问。它不为生成静态图表，而是建立可供持续探索的结构化知识库，支持 Claude Code、Cursor、Codex 等主流 AI 编码工具。项目通过解析代码结构、构建语义连接，把离散文件的逻辑关系具象化，帮助团队加速熟悉遗留系统、定位业务逻辑或在复杂代码库中导航。适合需要快速理解大型或陌生代码库的软件工程师。
Understand Anything is a developer tool that transforms any codebase into an interactive, queryable knowledge graph. It's designed for developers who need to quickly onboard onto unfamiliar projects, debug across module boundaries, or answer architectural questions without manually tracing thousands of files. The project positions itself as a learning accelerator—"Graphs that teach > graphs that impress."
Understand Anything 是一个开发者工具，它能把任意代码库转换成一个可交互、可查询的知识图谱。它的目标用户是那些需要快速熟悉陌生项目、跨模块排查问题，或者不想手动翻几千个文件就能摸清架构的工程师。项目的理念是“能帮你学懂的图谱，胜过只能看的酷炫图”。
Navigating a large legacy codebase or even a fresh microservice repo often means reading files linearly, jumping between definitions, and holding a mental model in your head. Traditional static analysis tools produce visualizations that are hard to interact with. Understand Anything addresses this by creating an explorable graph where entities (files, classes, functions) are nodes and their relationships are edges. You can search with semantics and ask natural language questions directly on the graph.
面对一个大型老旧系统或一个新接手的微服务仓库，通常只能线性阅读源码、跳来跳去查定义，脑子里拼命维护一堆关系。传统静态分析工具产出的可视化图往往无法交互。Understand Anything 解决了这个问题：它生成一个可探索的图谱——把文件、类、函数作为节点，它们之间的关系作为边——支持语义搜索，还能直接在图上用自然语言提问。
The project focuses on three key actions: building the graph, exploring it interactively, and integrating with your existing AI coding assistants. Notably, it prioritizes an interactive web UI over static export, making exploration feel like navigating a dynamic map. It also ships as a set of agent skills rather than a standalone monolithic app, allowing it to plug into Claude Code, Codex, Cursor, Copilot, Gemini CLI, and other editors seamlessly.
项目的核心能力围绕三件事展开：构建图谱、交互式探索、以及接入你现有的 AI 编程助手。它很重视交互式的 Web 界面（而不是只导出一张静态图），让探索就像看动态地图。另外，它被设计成一系列 agent 技能，可以无缝接入 Claude Code、Codex、Cursor、Copilot、Gemini CLI 等编程助手，而不是一个独立的大型单体应用。
The repository structure reveals a modular, agent-driven pipeline. The /skills directory contains dedicated skill definitions for different analysis tasks (e.g., codebase indexing, relationship classification, graph projection). These skills are consumed by an orchestrator agent (likely defined in the /api or /src layer) which coordinates code scanning, LLM calls for entity extraction, and graph construction. The built graph is then served through a web dashboard (in /public and /views) where you can browse, search, and ask questions.
从仓库结构能看出一个模块化的、agent 驱动的流水线。/skills 目录里放的是针对不同分析任务定义的技能（比如代码库索引、关系分类、图谱投射）。这些技能被调度 agent（很可能定义在 /api 或 /src 层）调用，来协调代码扫描、大模型提取实体和图谱构建。最终构建好的知识图谱通过 Web 看板（/public 与 /views）呈现，供你浏览、搜索和提问。
Although the README emphasizes an interactive experience from the official website, local usage revolves around installing the agent skills for your editor. The core invocation pattern follows a simple semantic command. For example, in Claude Code you would specify the skill and point it to your repo:
--skill codebase-knowledge-graph --path ./
This triggers the agent to analyze the current directory and provide an interactive graph link in the chat. For direct API exploration, the project serves a local dashboard on a designated port after running the build scripts outlined in the repository.
虽然 README 强调通过官方网站进行交互，但本地使用主要是在你的编辑器中安装 agent 技能。调用的核心模式是一个简单的语义命令。比如在 Claude Code 里，指定技能并指向你的仓库即可：
--skill codebase-knowledge-graph --path ./
这会触发 agent 分析当前目录，并在聊天里返回一个交互式图谱的链接。如果想通过 API 直接探索，项目在本地构建脚本运行后会在指定端口启动一个看板。
Ideal scenarios include onboarding onto large legacy systems, auditing technical debt, building a persistent project memory for AI agents, or visualizing a decoupling plan during architecture reviews. Boundaries and limitations: Graph quality heavily depends on the underlying LLM's code comprehension—it may hallucinate with deeply tangled global variables or heavy macros; the focus is primarily on static code-level entity relationships, with limited modeling of runtime behaviors like dynamic injection or reflection; initial indexing of extremely large monorepos can incur significant time and token costs.
适用场景：新人接手大型祖传代码、技术债梳理、为 AI agent 构建项目记忆、或是架构评审时可视化解耦方案。边界与限制：图谱质量高度依赖底层 LLM 的代码理解能力，面对极度混乱的全局变量引用或宏魔法时代码可能产生幻觉；目前侧重代码层面的实体关系，对运行时行为（如动态注入、反射）的建模有限；对于超大单体仓库（monorepo），初次索引的时间和 token 成本可能较高。