chat-widget - Drop-in AI chat widget pointed at your own gateway endpoint

A single-file AI chat widget you drop onto your own website. It renders a floating bubble, streams answers token-by-token from an Ollama-compatible endpoint, formats links/e-mails/phone numbers, offers quick-reply chips, and goes fullscreen on phones. No build step, no framework, no external CDN, no tracking, no cookies.

What you receive

Item	File
Widget source (HTML + CSS + JS in one file)	`src/widget.html`
Configuration template with every option documented	`src/config.example.js`
This README + license + changelog	`README.md`, `LICENSE`, `CHANGELOG.md`
Setup guide, embed snippet, troubleshooting	`docs/USAGE.md`
30 days of setup help by e-mail after purchase	included with the Setup Package

You do not receive a hosted service, a model, or an API key. The widget is a front end: it calls an endpoint that you own and operate. That is deliberate — your conversations and your key stay on your infrastructure.

Requirements

Any static web host (or an existing site — the widget is a fragment you paste in).
An Ollama-compatible endpoint reachable over HTTPS that accepts POST {"model": "...", "prompt": "...", "stream": true} and returns newline-delimited JSON with a response field.
That endpoint must allow your website's origin (CORS). See docs/USAGE.md → Troubleshooting.

Install in 3 steps

Copy the files. Put src/widget.html and a copy of src/config.example.js renamed to config.js in the same directory on your server.
Edit config.js. Set endpoint to your gateway (for example https://YOUR-HOST/ollama), set model, title, accentColor, welcomeMessage and systemPrompt. Never put an API key in this file.
Open or embed. Visit the page, or paste the widget block from src/widget.html plus the two <script> tags into your existing template. Done.

Full instructions: docs/USAGE.md.

Configuration at a glance

Key	Default	Purpose
`endpoint`	`''` (required)	Your gateway base URL; `/api/generate` is appended
`model`	`llama3.2:latest`	Model name as your backend knows it
`title`	`Chat with us`	Header text
`accentColor`	`#00e5a0`	Brand colour; readable text colour is chosen automatically
`welcomeMessage`	`Hi! How can I help you today?`	First message in the transcript
`inputPlaceholder`	`Type your message…`	Input hint
`systemPrompt`	generic assistant	The assistant's instructions
`promptUrl` / `promptVersion`	`''`	Optional: load the prompt from a text URL
`chips`	`[]`	Quick-reply buttons `[label, message]`
`position`	`right`	`right` or `left`
`statusLine`	`true`	Show the `Model: …` line
`idleTimeoutMs`	`60000`	Abort if the gateway sends nothing for this long

Known limitations (be aware before you buy)

There is no conversation history: each message is sent with the system prompt plus that one question. The model does not see earlier turns.
There is no server component in this package. You supply the endpoint; the key lives there, not in the browser.
There is no persistence: nothing is written to localStorage or cookies, so the transcript is gone on reload (and there is no cookie banner to add).
Markdown support is intentionally minimal: bold, italic, inline code, links, line breaks.
The widget is an inline <style> + <script> block, so a host page with a strict Content-Security-Policy must allow those blocks (hash/nonce) or serve them as files from its own origin. It uses no eval, no CDN and no third-party requests.

License

Commercial license — see LICENSE. Use it on sites you own or operate; you may not resell or redistribute the widget itself as a product.

Support

30 days of setup help are included with the Setup Package. Reply to your purchase receipt with the URL where you installed the widget and the exact error text (browser console + network tab) — that is usually enough to diagnose CORS and endpoint issues in one round trip.