OpenClaw is an open-source self-hosted AI assistant. This repository is an independent reimplementation of it in a different stack — FastAPI, LangChain and LangGraph on the server, Next.js on the client — written so the mechanisms are legible instead of hidden behind a framework.

It rebuilds the parts that matter. One agent is assembled from seven tools: a sandboxed terminal, a Python REPL, URL fetch, a file reader, knowledge-base search, browser automation and web search. Skills live as `skills/<name>/SKILL.md`, a scanner compiles them into a snapshot for the system prompt, and the agent has to read a skill file before it may use that skill. Memory works the same way — SOUL.md, IDENTITY.md, USER.md, AGENTS.md and MEMORY.md are fed in order, and RAG mode retrieves memory chunks per question instead of injecting the file whole.

Nothing is hidden while it runs. The chat streams tokens, tool calls, retrieval hits and raw messages as SSE events, and canvas HTML the agent emits previews in a sandboxed panel. Sessions persist as JSON with automatic titles, tiktoken counts and history compressed into a summary. It is deliberately single-user: CORS is open and there is no authentication, because it is built to run on your own machine.

Read it to learn how an agent works end to end — which is the entire point of rebuilding one rather than reading the original. The token stream, the tool registry, the skill scanner and the memory loader are each small enough to follow, and each one is the mechanism behind a behaviour you have probably only seen as a black box. OpenClaw itself belongs to its own authors; this is an independent implementation, not affiliated with or endorsed by that project.