How to Deploy AI-Built Websites in 2026: The Complete Guide

AI tools have become very good at making websites. Lovable, Bolt.new and v0 build whole apps in a browser tab, Claude and ChatGPT write a page in a single reply, and Claude Code, Codex and Cursor write projects straight to your disk. Most of them show you a preview, and several will host the result on a subdomain of their own. The questions start when you want your own domain, a host you choose, or a site that keeps running when you stop paying for the builder.

This guide is about that last step, as it stands in October 2026. It covers the three ways to get an AI-built site online, the route for apps that need a server, a decision tree by output type, and a matrix of 25 AI tools: what each one gives you, how you get it out, whether its own free plan can use your domain, and the shortest route from there. Tool facts were checked against each vendor's own documentation on 13 October 2026. Where a vendor's docs did not settle a question, the matrix says "check".

What a deployed AI-built site needs

A preview is not a deployment. Previews run in a sandbox or a development server, often with hot reload and other shortcuts that only exist while you build. A deployed site needs five things: a public address, HTTPS, your own domain once people outside your team will see it, a server that stays up after you close your laptop, and a way back when an AI regeneration breaks something.

What you hand to a host comes in four shapes:

• Plain HTML files. Often a single index.html, sometimes with a stylesheet and a script next to it. Many chat artifacts and DeepSite projects look like this. Nothing to build.
• A project that needs a build. A folder with a package.json: Vite, Astro, Next.js, TanStack Start. The build turns it into the files a browser loads, usually in dist/, build/ or out/.
• A Git repository. Where app builders such as Lovable, Bolt.new and v0 sync your code, and where coding agents already work.
• An app with a server. Its own API, server-side rendering or a database connection. Files alone do not run it.

The shape decides the route, so it pays to know which one you have before you pick a host.

Three ways to deploy, plus one for backends

1. Your AI deploys it over MCP

MCP (Model Context Protocol) lets an AI client call tools on outside services. VibeDeploy runs an MCP server at https://mcp.vibedeploy.be/mcp with 40+ tools: deploy a site, edit live files, connect a domain, switch on the forms relay, read visitor analytics. You add the server once and sign in through OAuth. From then on, "deploy this site to VibeDeploy" is the whole deploy step: the AI uploads the files and replies with the live URL, and later chats can edit and redeploy the same site. The mechanics are explained in What is MCP hosting.

Adding the server is one step in each client:

• Claude: on claude.ai or in Claude Desktop, open Customize > Connectors, click + Add, choose Add custom connector and paste the URL (if Claude asks for an OAuth client, choose Register automatically). The connector then works on the web, in Desktop, on mobile and in Cowork. Claude's help center says the Free plan includes one custom connector. The connector walkthrough has every screen.
• Claude Code: run claude mcp add --transport http --scope user vibedeploy https://mcp.vibedeploy.be/mcp, then /mcp inside a session to sign in.
• Codex CLI: run codex mcp add vibedeploy --url https://mcp.vibedeploy.be/mcp, then codex mcp login vibedeploy.
• Cursor, VS Code with Copilot, Windsurf (now Devin Desktop), Gemini CLI and ChatGPT (as a custom MCP server, on plans with write access): install links and config snippets are on the MCP setup page.

One thing the MCP server does not do is restore an old version. Every deploy is kept as a version with a restore point taken before it, and bringing an earlier one back is one click in the site's History tab in the dashboard, within your plan's restore window: 3 days on the trial, Solo and Maker, 7 on Studio, 30 on Business.

2. Paste the deploy guide into any chat AI

Not every assistant can add an MCP server. For those, the dashboard has a deploy guide (labelled Magic File there): one text file with instructions and a private deploy key. Paste it into ChatGPT, Gemini, Grok, Mistral or any other chat. If the AI can make web requests, it sends the files to VibeDeploy's deploy API itself and answers with the URL. If its sandbox has no network, it gives you a deploy page to open in your browser instead, with download and copy buttons as a fallback. Because the guide carries your deploy key, keep it out of public files and out of Git.

3. Upload a build or connect GitHub

The dashboard takes a ZIP, a folder or a single HTML file. A lone HTML file is served as the site's index.html, and a ZIP whose files sit inside one top-level folder is unpacked from that folder. A project that still needs its build, meaning a package.json without build output, is rejected with UNBUILT_PROJECT. Run the build first and upload what it produces:

npm install
npm run build
# upload dist/ (Vite) or build/ or out/ (Next.js static export)

If the code lives on GitHub, connect the repository instead. Every push to the branch you pick starts a build in an isolated build pod: npm ci (or npm install when there is no lockfile), then npm run build, and the first of dist, build, out or public goes live. Builds follow pushes to that one branch; there are no pull-request previews.

Backends: containers next to the site

Static sites and single-page apps are served directly. A site can also run containers next to it (your own server, Postgres, MySQL, Redis or workers; up to 6 services with 2 CPU and 2 GiB in total), deployed by your AI through the MCP tool deploy_app and isolated with gVisor. Container data is not yet included in snapshots.

A few practical rules come with that. Containers can reach the internet, but not private networks and not SMTP port 25. API keys and OAuth client secrets go into the dashboard, under the site's Settings tab in the Secrets section, and never into the chat. There is no dashboard view of container logs yet; your AI reads them with get_app_status. This is the route for a Node or Python API, a server-rendered framework, or a database your front end talks to.

Where to host a vibe-coded app: decide by output

Vibe coding leaves you with one of five things. Here is how each one gets online, and the trap that goes with it.

Static output is the easy case. VibeDeploy serves clean URLs, so /about is answered by about.html, which is the file layout a Next.js static export produces. Single-page apps get the fallback to index.html they need for client-side routes. A site with a top-level 404.html returns real 404s instead.

Server-rendered output is growing. Lovable apps created since 13 May 2026 run on TanStack Start, and Lovable's own hosting guide says such an app needs a host that runs a server, not one that only serves files; older Lovable projects on React + Vite still build to static files in dist/. Next.js lists the features a static export cannot use, among them cookies, redirects, Server Actions and incremental static regeneration. Google AI Studio apps can include a server-side Node.js component. All of these need running code. A host that only serves files cannot run them; on VibeDeploy the server part runs as a container next to the site.

Internal tools. For a tool only your team should see, make the site password-protected, private or invite-only (Maker and up). That protection is enforced by the server that delivers the static files, so it covers the static part. A tool with its own server in a container can sit behind an IP allowlist (also Maker and up) or handle sign-in itself.

When to stay where you are. If an app leans on one platform's own runtime, such as its edge functions or a builder's built-in database and auth, moving only the front end can leave you maintaining two places. Sometimes the honest answer is to keep it on that platform and upgrade for your domain there.

The tool matrix: 25 AI tools and how their output gets online

Here is how to read it. "What you get" is the form the output takes. "Getting it out" is the documented export. "Your domain on its free plan" asks whether the tool's own free plan can put the site on your domain, per the vendor's docs. "MCP" says whether the tool can add a remote MCP server such as VibeDeploy's: "yes" means the install steps are on our MCP setup page, "untested here" means the vendor documents MCP support but we have not run a deploy through it. "Check" means the vendor's documentation did not answer the question on 13 October 2026.

Chat assistants

App builders

Coding agents and editors

What the matrix shows

• Your own domain is mostly a paid feature on the builders. Lovable, Bolt.new, Bubble, Base44 and Anything put custom domains on paid plans. v0 is the exception, because it publishes to Vercel and Vercel's Hobby plan accepts custom domains, for personal, non-commercial use only.
• GitHub sync is the cleanest way out. Lovable, Bolt.new, v0, Replit and Google AI Studio all keep a repository in sync. Connect that repository to a host and every push rebuilds the site, while you keep building in the tool.
• Most coding agents and editors bring no production hosting of their own, and that is where MCP fits. Windsurf's App Deploys give a windsurf.build preview address, the hosting built into Codex is ChatGPT Sites (in the ChatGPT desktop app, on Plus, Pro and workspace plans), and Claude Code's artifacts are single pages on claude.ai. Claude Code, Codex, Cursor, VS Code and Windsurf each add a remote MCP server in one step, after which the agent that wrote the code also ships it.
• More output now needs a server. New Lovable apps, Next.js apps that use server features and AI Studio apps with a server part all need a host that runs code, not only files. That is the dividing line between a static host and a host with containers.

Your own domain and HTTPS

Every VibeDeploy site gets a vibedeploy.be address straight away, so you can share a working link while the domain is still being set up. To use your own domain, add it to the site in the dashboard, or ask your AI, which uses the tools add_custom_domain and verify_custom_domain. Then publish the DNS records you are shown: a TXT record that proves the domain is yours, and a CNAME to proxy.vibedeploy.be or, for an apex domain, an A record. Once DNS points at VibeDeploy, the certificate is issued automatically through Let's Encrypt. The trial includes one domain: link one you already own, or buy one in the dashboard.

Mistakes that break AI-built deploys

• Treating the preview as the site. Previews lean on development servers and sandbox tricks that do not survive a build. Build locally and open the output before you ship it.
• Hardcoded local addresses. AI tools like to leave http://localhost:3000 or 127.0.0.1 in fetch calls. Search the code for both before deploying.
• Secrets in front-end code. Vite puts variables prefixed with VITE_ into the client bundle, and its docs warn they should not contain sensitive information such as API keys. Anything in the browser bundle is public. Keep keys on a server: in a container whose secrets you set in the dashboard, or in your backend service.
• Uploading source instead of the build. A src/ folder full of TypeScript does not run in a browser. Upload dist/, or connect GitHub and let the build run for you.
• No 404 page on a multi-page site. Without a top-level 404.html, a mistyped URL shows your home page instead of an error.
• No way back. A bad regeneration can reach production in seconds. On VibeDeploy every deploy is kept as a version with a restore point before it, so the previous version is one click away in the History tab.

What it costs on VibeDeploy

The trial runs 14 days without a card: one site, and one custom domain you already own can be linked. After that, Solo is EUR 4 a month for one site and one domain, Maker EUR 15 for three sites with private, invite-only and IP-allowlist access, and Studio EUR 39 for ten sites with staging. Prices include VAT where it applies. There is no bandwidth overage billing; when a site outgrows its plan, you upgrade. If the trial ends without a plan, the site goes offline and your account and files are kept for 30 days, so choosing a plan brings it back at the same address. Details are on the pricing page.

Sites and their containers run on VibeDeploy's own Kubernetes cluster in Belgium, in one region and without a global edge network, under EU jurisdiction. The eight subprocessors are listed on the subprocessors page, and customer site content does not pass a US processor unless you use the optional AI builder features. If GDPR is part of your brief, the GDPR guide for AI-built sites covers what the regulation asks of a host.

Guides for each situation

• Localhost to a live URL: a project running on your machine, a temporary tunnel or permanent hosting.
• Download a Bolt.new project and deploy it: the ZIP, the build and a production URL.
• Put a Lovable site on your own domain: React + Vite and TanStack Start projects, step by step.
• Host a Claude artifact on a custom domain: download the artifact, host it, connect the domain.
• Claude saved your site to /home/workdir/artifacts/index.html: that path is in Claude's sandbox; here is how to get the file out.
• Deploy from Claude with one custom connector: Claude on the web, Desktop, mobile and Cowork.
• Publish a Claude Design website: from the export to your own domain.
• Move a Bolt.new project to Claude Code: keep the code, keep the site live.
• What MCP hosting is: how an AI agent deploys through tools.
• GDPR and AI-built sites in Europe: what a host has to give you.
• Claude Code from the terminal to a live site, what vibe deploying means and a landing page with a working signup form.

Sources

VibeDeploy facts come from our own product and documentation. Third-party facts were checked on 13 October 2026 on the vendor's own pages:

• Anthropic: custom connectors using remote MCP (plans, the Free plan limit, the Customize > Connectors steps); what artifacts are and sharing artifacts (copy and download, who can open a share); Get started with Claude Design (export formats, Handoff to Claude Code); Claude Code overview, MCP and artifacts.
• OpenAI: Creating and using ChatGPT Sites (plans, creating Sites in Work or in Codex in the desktop app, custom domains, production URLs); the ChatGPT Sites developer guide; Codex MCP and cloud tasks.
• Google: Gemini Canvas (Share & export, share links on gemini.google.com); AI Studio deploying, Build mode and full-stack apps; Gemini CLI MCP servers.
• xAI: Grok Build on web and mobile (grok.me, custom domains, GitHub) and Grok connectors.
• Mistral: Mistral Vibe, formerly Le Chat and MCP connectors.
• Lovable: custom domains, GitHub sync, FAQ (TanStack Start since 13 May 2026), hosting outside Lovable and custom MCP servers.
• Bolt: projects and download, GitHub, publishing, domains and MCP connectors.
• v0 and Vercel: v0 deployments, GitHub and FAQ; Vercel, adding a custom domain (domain limit for Hobby teams) and pricing (Hobby is for personal, non-commercial use).
• Replit: Starter plan, version control and MCP.
• Base44: quick start (ZIP and GitHub export) and billing and plans.
• Anything: publishing, export, backend and databases.
• DeepSite: the DeepSite Space and its source code (projects saved as static Spaces, ZIP download) and Hugging Face, Spaces custom domain.
• HeyBoss: FAQ. Softgen: FAQ. Tempo: documentation index.
• Bubble: pricing and application and data ownership.
• Cursor: MCP. VS Code: MCP servers. Devin Desktop: the Windsurf rename and App Deploys. Trae: MCP overview.
• Frameworks: Next.js, static exports; Vite, env variables and modes.