Ask AI for your Docusaurus docs, answered by the reader's own AI (no API key, no server)

[jbpin]

Hi all 👋 I built a plugin for my own docs site and open-sourced it: @leuria/docusaurus (Apache-2.0).

The idea: your readers already pay for an AI (ChatGPT, Claude) or run one (Ollama, LM Studio). Let that AI answer from your docs. You need no API key, no backend, and no per-question bill.

1. Ask AI
An "Ask the docs" panel in the navbar. It answers from your pages with links to them, knows which page the reader has open and what text they selected, and can open the right page for them.

2. Search by meaning (search: true)
⌘K search that matches words immediately and meaning once the reader has an embedding model: their local one, or an optional ~40 MB in-page model downloaded only with their consent. The index lives in the reader's browser and never leaves the device.

3. Your content as tools
The assistant works through four docs tools: search_docs, read_page, list_pages, open_page. The same tools are offered to the browser's own agents via WebMCP, so agent browsers can use your docs properly instead of scraping them.

4. Skills
Ship how-to playbooks for the reader's AI in the Agent Skills format (SKILL.md): from your site (static/.well-known/agent-skills/) or a GitHub repo pinned to a commit. Readers see which skills the site gives, and they apply to your site's conversations only.
skills: ["/", "your-org/your-skills#"]

Install
npm install @leuria/docusaurus

// docusaurus.config.js
plugins: [["@leuria/docusaurus", { search: true, suggestions: ["How do I get started?"] }]]

Works next to Algolia DocSearch or local search; no theme component is replaced.

Readers without their own AI still get answers: the panel falls back to the browser's built-in model where there is one, or to an OpenAI-compatible endpoint you choose (server). Example: a LiteLLM gateway with one cheap model, rate-limited, and no key in the page.

Docs: https://leuria.dev/docs/developers/docusaurus · Repo: https://github.com/leur-ia/leuria

It's early (v0.1). I'd love feedback, especially from anyone running large or versioned docs. What would stop you from trying it?