DeepSite v2 – Free AI Website Builder (DeepSeek V3/R1)

you can use on Hugging Face - DeepSite v2

Overview

DeepSite v2 is an open-source AI website/app generator hosted on Hugging Face Spaces. It turns natural-language prompts or Markdown into complete HTML/CSS/JS with Tailwind, then lets you iterate via structured diffs—no boilerplate or starter kits needed.

Under the hood, DeepSite uses Hugging Face’s Inference Client with provider routing to popular DeepSeek models (V3 and R1). You can stick with the default routing or bring your own provider keys for speed/cost control.

Key Features & Benefits

• One-shot site generation → Ship a working, responsive site from a single prompt. Tailwind CSS and sensible UI libraries included.
• Multi-page builds → Ask for multiple pages at once; links are wired so users can navigate immediately.
• Safe, auditable edits → Update pages using SEARCH/REPLACE diff blocks—no full rewrites. Great for code reviews.
• Provider flexibility → Switch among Fireworks, Novita, Together, Groq, Nebius, SambaNova to balance cost, latency, and context size.
• Streaming output → See code as it’s generated, so you can test quickly.
• Local run → Clone and run with an HF token; ideal for private or heavier workloads.
• Project persistence → Optional MongoDB backend for saving projects in self-hosted setups.

How It Works — Step by Step

1. Open the demo and pick a model/provider (or leave on Auto).
2. Describe your site in plain English (or paste a Markdown sketch).
3. Generate to receive index.html plus any additional pages.
4. Iterate by asking for changes; DeepSite replies with precise SEARCH/REPLACE diffs.
5. Add pages using the NEW_PAGE format; nav links are updated automatically.
6. Run locally if you need full control or want to attach your own provider keys.

Use Cases & Audiences

• Indie hackers / SaaS MVPs: “B2B SaaS landing + pricing (3 tiers), features, and docs home. Minimalist Tailwind, blue/gray palette.”
• SEO & content teams: “Multi-page review hub for [your product], with comparison table and FAQ schema.”
• Agencies/freelancers: Rapid mockups for client approval; iterate via diffs to track scope.
• Education: Teach HTML/CSS by diffing AI changes against student code.

Example Outputs / Applications

• Marketing site: 4 pages (Home/Pricing/Features/Contact), hero section with Feather Icons and AOS scroll effects.
• Event microsite: Countdown timer, schedule grid, speaker cards, RSVP form (front-end only).
• Portfolio: Masonry gallery with tags; mobile-first; dark mode toggle.

Expected latency: a few seconds to tens of seconds depending on provider/model and output length. (Choose faster providers like Groq when available.)

Technical Snapshot

DeepSite is a Next.js app that streams responses from Hugging Face’s Inference Client. It prompts a DeepSeek-family model to emit site code with Tailwind and lightweight JS, and it enforces a strict output contract (TITLE/NEW_PAGE/UPDATE + SEARCH/REPLACE markers) so edits remain stable. Provider routing enables you to pay with included HF credits or your own provider keys. Projects can be saved to MongoDB if configured.

Pricing & Limits

• Free to try on the public HF Space; you may need to log in for continued use. Anonymous requests are limited in the code to 2 per IP before login.
• Credits & providers: If you rely on HF-routed tokens, usage draws from your included Inference credits; BYO provider keys bypass HF’s included-credit ceiling.
• License: App code is MIT; outputs are subject to the model/provider’s license/ToS.

Privacy & Safety

• Data: Auth uses cookies; project saves rely on a MongoDB connection you control when self-hosting. The public Space does not publish a separate privacy policy—treat prompts/code as potentially logged by the chosen provider.
• Limitations: Static front-end focus; external CDNs for icons/animations; ensure license-safe images (placeholders recommended).
• Tips: Avoid including secrets in prompts; export locally before adding API keys; pin provider/model versions for reproducibility.

Comparisons & Alternatives

• DeepSite vs Vercel v0: DeepSite is open-source and model/provider-flexible; v0 is tightly integrated with Vercel.
• DeepSite vs Replit/“bolt”-style agents: DeepSite enforces a strict diff contract ideal for safe edits; agents often rewrite entire files.
• DeepSite vs Framer AI: Framer is design-led with a paid SaaS; DeepSite outputs raw code you fully own.

FAQ

Is it really free?
Yes—there’s a free public Space. Heavy use may require login and/or your own provider keys; HF-routed use consumes included credits.

What models does it use?
DeepSeek-V3 and DeepSeek-R1; more may appear in the UI.

Can I use my own API keys?
Yes. Add keys for providers like Together, Fireworks, Novita, Groq, etc., in HF settings and/or environment variables when self-hosting.

How do I run it locally?
Clone the Space, set HF_TOKEN (and optionally DB env vars), then npm run build && npm run start.

Does it support multi-page sites?
Yes; it can generate several pages and update nav links automatically.

How do edits work?
You send follow-ups; it replies with SEARCH/REPLACE diff blocks so you only patch what changed.

Is there an API?
No public DeepSite API. It’s a web app that calls HF Inference under the hood.

Which languages are supported?
UI is in English; generation is multilingual (model-dependent).

What about backend features?
DeepSite focuses on front-end sites. You can integrate real APIs later; start with fetch-based stubs.

Troubleshooting “credits exceeded”? Attach your own provider key or switch providers; HF-routed usage may hit included-credit ceilings.