The model wrote it. Now put it online.
Getting a website out of a chat is the easy half — every model will write you one. The hard half is that the HTML lands in a message bubble, and a message bubble isn't an address. harvis is the step after: hand over the files and get a URL back.
Where people building with AI reach for this
HTML sitting in a chat window
Copy it into a file, drop the file in, and it's live. There's a guide per chat app, because they all put the copy button somewhere different and some of them mangle the code on the way out.
An agent that can finish the job itself
Connect the MCP server and your agent gets a deploy_site tool: it publishes the folder and reads the live URL back to you, with no browser involved.
Iterating on a page that's already up
Redeploying replaces the site in place and keeps the address, so "make the header smaller" produces another deploy rather than another link.
From a folder to a link
- 01
Get the files out of the chat
One index.html is enough. If the model split things across messages, ask it for a single self-contained file — that's a complete site.
- 02
Publish them
Drag the folder into your browser, or let the agent call deploy_site over MCP and skip the browser entirely.
- 03
Share what comes back
A live URL and a private claim link. The claim link attaches the site to a free account whenever you want to rename it or replace it.
Pick how you drive it
All of these publish the same way and produce the same site. The only difference is what you already have open.
- MCP sunucusuThe agent deploys directly — no account and no API key, and it reads the finished URL back into the conversation.
- Dosya bırakSave what the chat wrote as index.html and drag it in. This is the shortest path from a message to a link.
- HTML StudioPick an open-weights model, describe the site, edit the code it writes, and deploy from the same screen.
- AI oluşturucuThe builder here, when you'd rather describe the page than prompt for it somewhere else and carry the result over.
Put something online and see
The first deploy needs no account: drop a folder in and a live URL comes back, along with a private claim link if you decide to keep the site.
Put a site onlineQuestions from people building with AI
Does the agent need an account or an API key to deploy?
No. The MCP server's deploy_site tool publishes anonymously and returns the live URL plus a private claim link. An account only matters later, when you want to rename the site, replace it, or take it down.
Can a model read the instructions for deploying here by itself?
Yes — /llms.txt on this domain is written for exactly that. Point a model at it and it has the upload contract, the limits, and the shape of a deploy without you explaining any of it.
The page the model wrote calls an API. Will that still work?
If the API is hosted elsewhere and allows browser requests, yes — the page is served as-is and its JavaScript runs in the visitor's browser normally. What won't work is code the model expected to run on a server, because harvis serves files and doesn't execute PHP, Node, or Python.
Do I need an account?
Not for the first deploy. You get a live URL and a private claim link; the claim link attaches the site to a free account when you want to choose its subdomain, redeploy over it, or delete it.
What can harvis not host?
Anything that needs code running on a server: PHP, Node, Python, a database, or a page rendered at request time. harvis serves the static files you upload and collects submissions from forms on them. A JavaScript front end talking to an API hosted elsewhere works normally.
What does it cost?
Publishing is free — 500 files and 50 MB per site, no card, and sites that do not expire. If a month outgrows the free limits the account steps onto the nearest paid step, from $5, and steps back down when usage does. Bandwidth and SSL are never billed.
Is someone else's job closer to yours?
The product underneath is the same — these pages just start from a different problem.