A chat widget that only knows your site.

It reads the pages you publish, answers from those, and says it doesn't know about anything else. Any model, your own API key, your own content. MIT.

npx askdock ingest https://your-site.com

<Askdock endpoint="/api/askdock" name="Acme AI" />

Most drop-in chat widgets are a general-purpose model wearing your logo. Ask one about a competitor's pricing and it will happily make something up — on your domain, in your voice.

Askdock is the other shape. The model gets no tools, no web access, and one closed knowledge base built from your own pages. What isn't in that file isn't available to it: not filtered, absent. When the site doesn't cover a question, it says so and points at a human.

   npx askdock ingest              your server                    the browser
┌────────────────────────┐   ┌──────────────────────┐   ┌──────────────────────┐
│  sitemap.xml           │   │  createAskdockHandler│   │  <Askdock /> or      │
│  → pages → plain text  │──▶│  corpus + rules      │◀──│  <script src=…>      │
│  → .askdock/corpus.json│   │  API KEY lives here  │   │  never sees a key    │
└────────────────────────┘   └──────────┬───────────┘   └──────────────────────┘
                                        │
                              Gemini · Claude · GPT · local

Three things keep it on topic, because any one alone leaks: no tools, a closed corpus, and a prompt that makes the refusal explicit. Every answer that used the corpus comes back with citations, checked against the real pages so a made-up id becomes nothing rather than a dead link.

npx askdock ingest https://your-site.com
#   24 pages · ~31.4k tokens · Your Site
#   → .askdock/corpus.json

It reads sitemap.xml, or follows same-origin links when there isn't one. The result is plain JSON you can open, read, diff and commit — the assistant's entire knowledge in one reviewable file. Re-run it whenever the site changes.

A folder of Markdown works too:

npx askdock ingest ./content --base-url https://your-site.com/docs/

// app/api/askdock/route.ts
import { createAskdockHandler } from "@askdock/server";
import corpus from "../../../.askdock/corpus.json";

export const POST = createAskdockHandler({
  model: { model: "google/gemini-3.1-flash-lite" },
  corpus,
  persona: { name: "Acme AI", fallbackHref: "/contact" },
});

export const maxDuration = 60;

The handler is a plain Request → Response function, so the same line works on Hono, Bun, Deno, or anything else with fetch.

React

npm i @askdock/react @askdock/server @ai-sdk/google

import { Askdock } from "@askdock/react";

<Askdock endpoint="/api/askdock" name="Acme AI" />

Anything else — Webflow, WordPress, Astro, plain HTML. One tag, ~14 kB gzipped:

<script
  src="https://unpkg.com/@askdock/embed/dist/askdock.js"
  data-endpoint="https://acme.com/api/askdock"
  data-name="Acme AI"
  defer
></script>

model: { model: "google/gemini-3.1-flash-lite" }
model: { model: "anthropic/claude-sonnet-4-5" }
model: { model: "openai/gpt-5" }
model: { provider: "openai-compatible", model: "llama-3.3-70b", baseURL: "http://localhost:11434/v1" }
model: { provider: "gateway", model: "anthropic/claude-sonnet-4-5" }   // Vercel AI Gateway

The prefix picks the provider package — install only the one you use: @ai-sdk/google, @ai-sdk/anthropic, @ai-sdk/openai. Pass an already-built AI SDK model if you need a provider this list has never heard of.

Set apiKey explicitly if your keys live somewhere else — a secret manager, a per-tenant lookup.

Pin the version, don't track an alias. -latest moves when the provider decides it does, and can land you on a tier with a different free quota without a commit of yours.

Launcher and surface are independent: a pill at the bottom can open a side panel, a corner bubble can open a popup.

Every colour is one flat set of custom properties, written on the widget root — so the widget can't repaint your site and your reset can't flatten the widget. The default is black and white on purpose.

The three easy to confuse: accent draws lines, brand fills shapes, subtle is a hover — and none of them is the visitor's message bubble.

Every option above that is a string has a data- twin. Appearance and theme keys are flat.

Anything an attribute can't express goes through window.askdock({ … }), which re-renders in place.

askdock ingest <url|directory> [options]

Nothing in the CLI reads or writes an API key.

A corpus from your CMS, instead of a crawl:

corpus: async () => ({
  site: { name: "Acme", url: "https://acme.com" },
  docs: (await cms.getPages()).map((p) => ({
    id: p.slug, title: p.title, url: `/${p.slug}`, kind: "Page", text: p.body,
  })),
});

Your own visual language, when the tokens aren't enough:

<Askdock
  endpoint="/api/askdock"
  icon={<LogoMark />}
  renderLauncher={({ open }) => <MyPill onClick={open} />}
  renderLink={(props) => <Link {...props} />}
/>

A full page instead of a floating panel:

<Askdock endpoint="/api/askdock" appearance={{ surface: "inline", launcher: "inline" }} />

Your own UI, with the conversation handed to you:

const { messages, streaming, error, ask, stop } = useAskdock({ endpoint: "/api/askdock" });

• The key is read on your server and sent to your provider. Nothing else.
• The widget knows one thing about your setup: the URL it posts to.
• There is no mode in which a key reaches the browser, and there won't be one — a key in a <script> tag is a key in everyone's browser.
• No commit in this repository has ever contained a key. .env* is ignored; .env.example ships empty.

See SECURITY.md for hardening a public deployment.

pnpm install
pnpm build
cp apps/demo/.env.example apps/demo/.env.local   # add a key
pnpm dev                                          # localhost:3000
pnpm test

• Models paraphrase, and paraphrase drifts. Grounding makes an unsupported answer unlikely, not impossible. Keep the disclaimer; point people at a human for anything that matters.
• Small models are the ones that break the rules — answering in the wrong language, declining what the corpus plainly covers. If the assistant misbehaves, try a larger model before rewriting the prompt.
• The bundled rate limit is in-memory and per instance.
• Retrieval past the prompt budget is lexical, not semantic. Big sites should pass their own selectDocs.

Pull requests are welcome and are the only way in: main is protected, and every change is reviewed before it lands. Start with CONTRIBUTING.md.

MIT — see LICENSE.