PandaDesk Lite
An AI support chat for one website, running on your own Cloudflare account with your own keys.
It reads your website and docs and answers visitors' questions from them. When it can't answer, or someone asks for a person, it takes their email and passes the chat to you. You reply from a small admin page.
- Chat widget: one script tag, your colour, greeting and starter questions.
- Answers from your content: it reads your website (using the sitemap when there is one) every day. You can also paste in text and pin answers to common questions.
- Search by meaning and by keyword, with links to the pages each answer came from.
- Hand-off to you: the visitor leaves their email, you're emailed, and you reply from
/admin. They see your reply in the chat and by email.
- Your keys: any OpenAI-compatible API (OpenRouter, OpenAI, Anthropic, Groq…), or Workers AI with no key at all.
It's built for one site with one admin. For a team inbox, an agent that looks up your customers' own data, several apps, or setup from your coding agent, see PandaDesk Cloud.
Deploy
Or from your machine:
git clone https://github.com/kamskans/pandadesk-lite && cd pandadesk-lite
npm install
npx wrangler login
# Search by meaning needs a Vectorize index; without it, search uses keywords only.
npx wrangler vectorize create pandadesk-lite --dimensions=1024 --metric=cosine
npx wrangler secret put ADMIN_PASSWORD
npx wrangler secret put LLM_API_KEY # optional: leave out to use Workers AI
npm run deploy
The Worker's home page describes PandaDesk Lite and has a live demo of your chat; change it in ui/src/home.tsx. Then open https://<your-worker>.workers.dev/admin, sign in, add your website under Knowledge, and copy the snippet from Install into your site:
<script src="https://<your-worker>.workers.dev/widget.js" async></script>
Settings
Secrets (npx wrangler secret put NAME, or .dev.vars locally; see .dev.vars.example):
| Name | |
|---|
ADMIN_PASSWORD | Required. Password for /admin. Changing it signs you out everywhere. |
LLM_API_KEY | Key for LLM_BASE_URL. Without it, answers use Workers AI (Llama 3.3 70B). |
RESEND_API_KEY, EMAIL_FROM | Optional. Send email through Resend: your replies to visitors, and alerts to you. |
NOTIFY_EMAIL | Optional. Where to email you when someone is waiting for a reply. |
Variables (in wrangler.jsonc):
| Name | Default | |
|---|
LLM_BASE_URL | https://openrouter.ai/api/v1 | Any OpenAI-compatible API. |
LLM_MODEL | openai/gpt-4o-mini | The model name for that API. |
SITE_NAME | Support | Default chat title. You can change it in the admin. |
MAX_PAGES | 300 | Most pages read per website. |
How it works
One Cloudflare Worker serves the home page, the widget, the chat, the admin page and the API. The pages are React and Tailwind, built from ui/ into dist/ with Vite before every wrangler dev and wrangler deploy.
- Data: conversations and knowledge are stored in D1. Passages are indexed in D1's full-text search and in Vectorize.
- Embeddings: made with Workers AI (
@cf/baai/bge-m3, multilingual).
- Background work: a cron trigger runs every 10 minutes. It reads pages that are waiting and re-reads each website once a day. While the Knowledge page is open, pages are read straight away.
On the Workers Free plan, small sites work fine. Large sites read faster on Workers Paid.
Develop
npm install
cp .dev.vars.example .dev.vars # set ADMIN_PASSWORD
npm run dev # builds ui/, then http://localhost:8787/admin
npm test && npm run typecheck
Workers AI and Vectorize run on Cloudflare even in local development, so wrangler dev asks you to sign in for them. Leaving them out of wrangler.jsonc locally falls back to keyword search. Set LLM_API_KEY to answer without Workers AI.
License
AGPL-3.0. You can use, change and self-host this freely. If you run a changed version as a service for others, you must publish your changes under the same license.
"PandaDesk" and the PandaDesk logo are trademarks of Kamal Kannan and aren't covered by the code license. If you publish a fork, please give it a different name.
