A tiny, anonymous chat page for DeepSeek V4 Pro 0813 with reasoning enabled at maximum effort. There are no accounts, registrations, app-set cookies, analytics, or database records. Conversation history lives only in the open browser tab.
The application source is licensed under AGPL-3.0-or-later. It uses a vanilla browser client, one small server function, native fetch, and a single direct development dependency: MIT-licensed Vite.
The server—not the browser—fixes the DeepSeek request to:
{
"model": "deepseek-v4-pro",
"thinking": { "type": "enabled" },
"reasoning_effort": "max",
"stream": true,
"max_tokens": 8192
}The August 13, 2026 DeepSeek release promoted DeepSeek-V4-Pro-0813 to GA behind the deepseek-v4-pro API alias. deepseek-v4-pro-0813 is not a public API identifier. See the official release notes, model list, and thinking-mode guide.
- Import this project into Bolt and run
npm installif Bolt has not already done so. - Ask Bolt to deploy
supabase/functions/chat/index.tsas the publicchatserver function. Keepverify_jwt = false; the app intentionally has no authentication. - Open Database → Secrets and create
DEEPSEEK_API_KEYwith your DeepSeek key. Never create aVITE_DEEPSEEK_API_KEY: variables prefixed withVITE_are public browser code. - Bolt normally injects
VITE_SUPABASE_URLwhen its database/server functions are connected. If it does not, setVITE_CHAT_ENDPOINTto the full public URL ending in/functions/v1/chat. - Publish the site. Once its URL is final, add
ALLOWED_ORIGIN=https://your-site.bolt.hostas a server-function secret and redeploy the function.
No auth tables, user accounts, or application database tables are needed. Bolt documents server-side secrets and functions in its Secrets and Server functions guides.
Requires Node 20.19 or newer.
npm ci
cp .env.example .env.local
# Put your key in .env.local, then:
npm run devOpen http://localhost:5173. .env.local is ignored by Git.
For the portable production server:
npm run build
DEEPSEEK_API_KEY=your-key npm run serveThe production Node server has no package dependencies: it serves dist/ and proxies /api/chat. Vite is loaded only by the development command. This makes the app self-hostable without Bolt or Supabase.
- The API key exists only in the server environment.
- The server accepts only
userandassistanttext, then adds its own system prompt. - The model, upstream URL, reasoning mode, effort, and output cap cannot be changed by visitors.
- Request bytes, message count, per-message size, total history, generation time, and request frequency are bounded.
- Model output is inserted as text, never as HTML.
- Prompts are not logged or stored by this application. They are sent to DeepSeek and are subject to DeepSeek's own policies.
The included rate limiter is only best-effort because serverless instances do not share memory. A public, no-login AI proxy can spend your API balance. Treat a persistent gateway limit and spend cap as launch requirements. V4 Pro is priced above V4 Flash per token, so identical traffic costs more than it did before this upgrade. Before sharing widely, use a dedicated API key, keep a low prepaid balance, enable monitoring, and configure a persistent rate limit at Bolt, Supabase, or another reverse proxy. Origin checks and CORS are not substitutes for rate limiting.
The original application code is AGPLv3-or-later, Vite is MIT, and the app avoids remote fonts, SDKs, analytics, proprietary UI libraries, and CDNs. Every build exposes the exact corresponding application source at /source-code.txt so a deployed visitor can inspect, modify, and run it.
Two unavoidable boundaries remain:
- Bolt Cloud is a hosted proprietary service chosen for deployment. The included Node server keeps the app portable to a FLOSS host.
- DeepSeek publishes V4 Pro model weights under the MIT license, but
api.deepseek.comis a managed external service. The provider's service terms still apply.
See THIRD_PARTY_NOTICES.md and the full GNU AGPL license.
npm run checkThis runs the unit tests, production build, and corresponding-source generation.
see https://artificialanalysis.ai/ for cost, etc.