Zia Vsor SDK — AI skill for Claude Code

vsor

**Answers you can trace. Refusals you can trust.**

**[See a site vsor built →](https://vsor-demo.vercel.app)** — four markdown files, one command.

Point an AI assistant at your own files and get answers that name the file they came from — and a plain *"that isn't in here"* when the question falls outside them.

uvx vsor init my-sor  # a project in your own repo: markdown, config, git initialized
cd my-sor
vsor dev              # the live site on 127.0.0.1:3000, reloading as you write
vsor build            # → build/, ready to upload anywhere

What vsor builds

One folder of markdown. Two surfaces, built from it, that can never disagree:

• A website your team reads. Search, dark mode, quizzes and flashcards, self-hosted fonts, and no calls to anyone else's server — it works behind a firewall.
• A connector your AI answers from (MCP — the standard assistants like Claude use to reach outside tools). Every answer names the document it came from; a question your files do not cover is declined instead of guessed.

Every build also writes `build.lock.json` — a record of exactly which documents went in, from which commit, with which versions. When someone asks *why did it say that*, that file is how you find out.

Your side of it is a folder. Sub-folders become sections; a line of frontmatter sets the title:

knowledge/
  about.md
  thesis.md
  ecosystem/
    the-ecosystem-concept.md
instance.md                   ← what this deployment is
site/docusaurus.config.ts     ← title, navbar, colours — the parts that are yours

Here is the website surface, from a real build of real content — no theme to pick, no components to wire, no build config:


![A generated document page: sidebar built from the folder tree, breadcrumbs, reading time, copy-as-markdown, and an in-page table o