DeepSeek Chat — Official UI with Web Search, Thinking & Voice

Deploy the official open-source DeepSeek chat client (deepseek-ai/chat-demo, MIT) as one static Railway service. It renders DeepSeek's full App API feature set — server-side web search with citations, the deep-thinking panel, file upload, and voice input/read-aloud — in a React UI with Chinese and English localization and light/dark themes.

Your API key never leaves your browser. The service is a pure static file server (Caddy): no backend, no database, zero variables, nothing stored server-side. Your key lives in your browser's localStorage and goes straight to https://api.deepseek.com; conversation history lives in your browser's IndexedDB.

Feature	What you get
Server-side web search	DeepSeek runs the search for you — result cards with titles, snippets and page ages, no SERP keys or scraping infra
Deep thinking	Expandable reasoning panel; thinking + search combine into deep search
Citations	Inline citation links and hover cards on search-backed answers
Files	Paste, drag, or pick attachments; server-side OCR and validation
Voice	Voice input (ASR) and read-aloud replies (TTS) with live audio spectrum
Markdown stack	Syntax highlighting, KaTeX math, Mermaid diagrams, code blocks
History	Multiple conversations, branching, edit/regenerate, ZIP import/export
zh/en + themes	Simplified Chinese and English UI; light and dark; mobile layout

Note: other Railway listings named DeepSeek Harness deploy the DeepSeek coding agent behind a login — a different product. This template deploys the official chat client itself.

After deploying

Open your app URL.
Settings → API Key: paste a key from platform.deepseek.com.
Chat. Toggle thinking and search in the UI when you want them.

Hosting cost on Railway is a few MB of RAM for a static container. API usage is billed to your own DeepSeek account at DeepSeek's published rates (flash model roughly 0.15–0.30 dollars per million input tokens; search/fetch 0.75 dollars per 1,000 uses; off-peak discounts). See the DeepSeek pricing page.

Churn disclosure: the DeepSeek App API and this client are both days old (client repo created 2026-11-12). This template pins upstream at commit 5d5dd754ebc1948e977ddb880a9eb5aa25f37b4c (v0.1.1); upstream will move, and re-pinning is deliberate.

Deploy and Host

About Hosting

This template provisions exactly one service: a digest-pinned caddy:2-alpine container serving the official client's fully static build. The Dockerfile clones deepseek-ai/chat-demo at the pinned commit SHA, builds it with pnpm 12.6.0 and rspack on node:22-alpine, and copies the resulting dist/ into a pure Caddy runtime — there is no Node process, no database, and no volume. Caddy serves the SPA with a fallback to index.html, zstd/gzip encoding, and security headers; the healthcheck probes the real /health route (200, unauthenticated, dotless). Deployments restart on failure up to 10 times. Since there is no server-side state, redeploys and restarts are free of migration or seeding concerns — all user state is in the visitor's browser (localStorage for the key and settings, IndexedDB for history and attachments).

Why Deploy

Official client, zero maintenance surface: you host a static site; the browser talks directly to the DeepSeek API. Nothing of yours ever sees an API key or a conversation.
Privacy pitch that is verifiable: the container contains nothing but static files. There are no variables, no volumes, no credentials, no logs of your chats — your key and history stay in your browser (clearing browser data clears history; export/import via built-in ZIP).
The whole new App API feature set out of the box: server-side web search with citations, deep thinking, file upload with OCR, voice input and read-aloud — features generic OpenAI-compatible chat UIs on Railway cannot render because they proxy the API themselves and skip DeepSeek's extensions.
Cheapest possible hosting: a pure Caddy static container is the smallest service shape on Railway, and API costs flow to your own DeepSeek account with off-peak discounts and cache pricing.

Common Use Cases

A personal DeepSeek chat UI on your own domain with web search, citations, and the thinking panel.
A private, key-your-own-device team portal: every teammate pastes their own key, keys and histories never leave their browsers, and the shared server holds nothing sensitive.
A mobile-friendly voice assistant: talk to DeepSeek, hear replies read aloud, and attach photos for OCR.
A research scratchpad: deep search answers with cited sources, Mermaid/KaTeX rendering, and ZIP history export for archiving.

Dependencies for

Deployment Dependencies

None at deploy time: the template has zero variables, zero plugins, and zero volumes. One static service, one auto-provisioned domain.
A DeepSeek API key from platform.deepseek.com, entered in the app's Settings after deploy (stored only in your browser). API usage is billed to your own DeepSeek account, not to Railway.
A modern browser: Chrome or Edge 105+, Firefox 110+, or Safari 16+ (per the client's own targets). Voice input additionally requires microphone permission.