Vibe Deploying: What It Means and How It Works

First there was vibe coding: describing software in plain language and letting an AI write it. The obvious next question was what happens to all that AI-written software when it needs to go live. The answer that has emerged has a name of its own.

What is vibe deploying?

Vibe deploying is shipping a website or app to a live URL by talking to an AI, instead of running a build pipeline, configuring a server, or clicking through a hosting dashboard yourself. You say "put this online" and the AI handles the upload and the hosting, then hands back a working HTTPS link. The term extends the logic of vibe coding one step further: if you never wrote the code by hand, there is no reason to deploy it by hand either.

That is the definition. The rest of this post covers where the line between coding and deploying sits, what actually happens under the hood when an AI ships a site, and where the honest limits are.

Vibe coding vs vibe deploying

Vibe coding is about creation. You describe a landing page, a portfolio, a small tool, and the AI writes the HTML, CSS, and JavaScript. The output is code, sitting in a chat window, a sandbox, or a folder on your machine.

Vibe deploying is about shipping. It picks up exactly where vibe coding stops: taking that code and putting it on the public internet, with a real address, a certificate, and a server that stays up after you close the laptop.

The distinction matters because many AI tools stop after the first half. ChatGPT, Claude, and their peers are excellent at producing a working site. Some will also host it on their own platform, such as ChatGPT Sites on Plus, Pro and workspace plans or a published artifact on claude.ai, but putting it on a host you choose, with a server or a database next to it when the app needs one, is still a separate job. Vibe deploying closes that gap. The coding half and the deploying half are done the same way, in the same conversation, by the same AI.

How a vibe deploy actually works

There is no magic in it, just two concrete mechanisms.

Mechanism 1: an MCP tool call. MCP (Model Context Protocol) lets you connect an AI client like Claude, Cursor, or Codex to outside services. Connect a hosting provider's MCP server once, and the AI gains deploy tools it can call directly. When you say "deploy this", the AI invokes the tool, transfers the files it generated, and reads back the live URL, all inside the conversation. This is the cleanest form of vibe deploying because no files ever pass through your hands. VibeDeploy's server lives at https://mcp.vibedeploy.be/mcp; the MCP setup page has the steps for each client, and What is MCP hosting explains the mechanics.

Mechanism 2: a pasted deploy guide. Not every AI tool speaks MCP. For those, the fallback is a deploy guide: a single instruction file you paste into the conversation. If the AI can make web requests, it reads the guide and performs the deploy through the hosting API on your behalf. If its sandbox has no network, it gives you a small deploy page to open in your browser instead, and the upload happens from there. Slightly less elegant, same end result: you talk, the site goes live.

Either way, the loop that follows is the good part. "Make the headline bigger" is followed by "deploy it again", and the same URL updates. Edit, ship, repeat, without leaving the chat. Tool-by-tool instructions for both mechanisms live in the deploy guides.

When vibe deployment is enough, and when it is not

Honesty time, because the term invites overreach.

Where it shines: static sites and single-page apps. Landing pages, portfolios, marketing sites, event pages, documentation, prototypes, and AI-built front ends, including ones that call external APIs or a backend service such as Supabase. These compile to files a host can serve as they are, which means an AI can take them from prompt to production with nothing else involved. Much of what people build with AI looks like this.

Where it now reaches further: backends. On VibeDeploy, an AI connected over MCP can also start containers next to the site with the tool deploy_app: your own server from the source tree, Postgres, MySQL, Redis or workers, up to 6 services with 2 CPU and 2 GiB of memory in total, isolated with gVisor. "This app needs a database, set one up next to it" is a request the AI can carry out, not a dead end.

Where you still have to think: the decisions around a backend. API keys and OAuth client secrets belong in the dashboard, in the site's Settings tab under Secrets, and never in the chat. Container data is not yet included in snapshots, so a database that matters needs its own export routine. There is no dashboard view of container logs yet; your AI reads them with get_app_status. An AI can set all of this up, but a person should still own the data and the secrets.

The practical takeaway: check which kind of project you have. If npm run build produces a folder of files, or your whole site is one HTML file, the deploy is one sentence. If the project needs a server or a database, it is still a conversation with your AI, just a longer one. The complete guide to deploying AI-built websites sorts 25 AI tools and every project type by route.

Try a vibe deploy yourself

The fastest way to understand vibe deploying is to do one:

1. Ask your AI of choice to build a small one-page site.
2. Connect VibeDeploy: add https://mcp.vibedeploy.be/mcp as an MCP server in Claude, Cursor, Codex or another client (steps per client), or paste the deploy guide from the dashboard into the chat.
3. Say "deploy this".
4. Open the live URL it replies with.

The whole loop, from an empty chat to a public HTTPS address, usually fits in a few minutes, and repeating a deploy after an edit is one more sentence. The 14-day trial needs no credit card and can link one custom domain you already own, so you can point a real domain at the result before deciding anything. Plans and limits are on the pricing page.

Once you have shipped a site by finishing a sentence, the old way, with the FTP client or the CI config or the dashboard upload form, starts to feel like the workaround.

Sources

• OpenAI Help Center, Creating and using ChatGPT Sites: Sites is in public beta for ChatGPT workspaces, Plus and Pro accounts, and is not available on Free or Go. Checked 13 October 2026.
• Anthropic Help Center, Share artifacts: only a legacy artifact published from a chat gets a public link that anyone can open without a Claude account; other shared artifacts need a Claude account to open. Checked 13 October 2026.