deni-ai-support

The Deni AI help center (support.deniai.app), built with Fumadocs on TanStack Start.

Development

pnpm install
pnpm dev

pnpm run types:check
pnpm run lint
pnpm run build

Writing articles

Articles live in content/docs/. Each article has an English file and a Japanese file next to it:

content/docs/flixa/external-tools.mdx     -> /en/flixa/external-tools
content/docs/flixa/external-tools.ja.mdx  -> /ja/flixa/external-tools

Sidebar order and folder titles come from meta.json (use meta.ja.json for a Japanese title that differs). / redirects to /en or /ja based on Accept-Language.

Support features

Every page has a "Get help" widget (bottom right) where an AI assistant (Claude Haiku 5.5) answers questions by searching and reading the help center articles. The contact and feedback forms, which email the team, also live in the widget: its "Contact support" and "Send feedback" buttons open them, and the assistant can draft either message as a card in the chat that the user reviews, edits, and sends. Messages sent from the widget carry the conversation. Links to /en/contact and /en/feedback (and the /ja/ versions, with an optional ?category= on contact) open the forms in the widget; visiting those URLs redirects to the help center home with the form open. Each article also ends with a "Was this page helpful?" prompt.

With a database configured, widget conversations are saved so the team can answer contact and feedback messages:

• Each chat exchange and each message sent to the team is stored in the support schema. The browser keeps the conversation's credential (<id>.<HMAC token>) in localStorage and restores the chat on the next visit.
• The team works from the admin inbox at /admin (password: SUPPORT_ADMIN_TOKEN), which lists conversations with a contact or feedback message. Notification emails to the team link to the conversation there.
• A reply appears in the visitor's chat (with an unread dot on the "Get help" button) and is emailed to the address they gave, with a link that opens the conversation in the widget on any device. The visitor answers with the reply's "Reply" button, which reopens the conversation in the inbox.
• "New chat" starts a new conversation; earlier ones stay reachable from the reply emails.

They are configured with runtime environment variables, listed in .env.example. For local development, copy it to .env.local. All are optional: without ANTHROPIC_API_KEY the widget shows only the contact and feedback buttons; without the Cloudflare email credentials the forms say they are unavailable; without both Turnstile keys nothing is challenged; without DATABASE_URL and SUPPORT_SECRET nothing is saved and replies go by email only. Requests are also rate limited per IP in memory.

Variable	Purpose
ANTHROPIC_API_KEY	AI support chat
CLOUDFLARE_ACCOUNT_ID, CLOUDFLARE_API_TOKEN	Contact and feedback email (Email Sending: Edit)
SUPPORT_EMAIL_TO, SUPPORT_EMAIL_FROM	Recipient and sender (default contact@deniai.app, Deni AI Support <noreply@deniai.app>)
TURNSTILE_SITE_KEY, TURNSTILE_SECRET_KEY	Cloudflare Turnstile for the chat and forms
DATABASE_URL	Postgres (the Deni AI database) for saved conversations; the pooled URL works
SUPPORT_SECRET	Long random string that signs conversation links and admin sessions; changing it detaches every saved conversation from its browser
SUPPORT_ADMIN_TOKEN	Password for the admin inbox at /admin

Database

The tables live in the support schema of the Deni AI database (src/server/db/schema.ts); deni-ai and Flixa leave that schema alone. After changing the schema, generate a migration and review the SQL in migrations/:

pnpm run db:generate

Apply migrations to the database in DATABASE_URL from .env (production) or .env.local (development) before deploying code that needs them:

pnpm run db:migrate

pnpm run db:migrate:dev

Deployment

Production runs on Dokploy from the Dockerfile (Build Type: Dockerfile, port 3000, domain support.deniai.app). Set the support variables above as runtime environment variables; no build arguments are needed. Run pnpm run db:migrate against the production database before the first deploy with DATABASE_URL.

pnpm run build produces a Node server in .output/. It still imports a few packages at runtime, so start it next to node_modules:

pnpm start

The server listens on PORT (default 3000).