Your AI made it. Now it has a link.
A page, a report, a folder of notes or a whole site — live on its own HTTPS address, straight from the chat. No repository, no build config, no dashboard in the loop.
Add it to Claude
One connector for claude.ai, the Desktop app and your phone.
Listed in Claude's connector directory · sign in with your Drop2Run account
$ claude mcp add drop2run -s user -- npx -y @drop2run/mcpOnly to publish folders and files already on your disk — a built site, a PDF.
Set it up in your agent
The same server in every client that speaks MCP. Add it, ask for a publish, and the first one opens a browser tab to sign in.
Cursor
EditorOpens Cursor with the server filled in. Press Install.
VS Code
Copilot agent modeAdds the server to your user profile.
Codex
CLI · IDERun once. It writes to
~/.codex/config.toml.$ codex mcp add drop2run -- npx -y @drop2run/mcpChatGPT
Developer modeAdd a custom connector with this URL. No developer mode? Paste the answer below.
https://mcp.dropto.run/mcpWindsurf
EditorIn MCP settings, add a server with this command.
$ npx -y @drop2run/mcpGemini CLI, Zed, Cline…
Any MCP clientTakes a command? Use this. Takes a URL?
https://mcp.dropto.run/mcp$ npx -y @drop2run/mcp
Config files, sign-in without a browser, and a token of your own:Full setup for each client →
From any chat: paste it, take the link.
HTML becomes a page; markdown becomes a document, opened in a reader with its tables, code and diagrams rendered. No account, no install, no file saved first — the same publish the tools make, from the answer you already have in front of you.
HTML is published as a page, exactly as pasted. Anything else is published as a markdown document and opened in the reader.
Publishing what Claude just wrote
The usual route is a download: open what the model produced, save the HTML, find the file, drag it somewhere that hosts it. It works, and it is four steps in three applications — which is three more places to lose a page that took one sentence to make.
Here there is no step outside the conversation. The model calls a tool with what it wrote and reads the address back to you, and the sign-in — the one part that normally sends somebody to a dashboard for a token — happens through the same tools. Nothing is saved to disk on the way unless you wanted it saved.
If you have never installed anything
Three steps, once. After that, publishing is something you ask for in a sentence.
- 1
Press one button
"Add from Claude's directory" at the top of this page. Claude asks you to connect, and that is the whole of it — no terminal, nothing to download, and it works on the Desktop app and your phone too.
- 2
Ask for a page
In the same conversation that wrote it: “publish this”. The model picks the tool itself; there is no command for you to remember and nothing to configure first.
- 3
Sign in where you are
The first publish opens a browser tab to approve, and that is the whole of it — no token to copy, no dashboard to visit. On a machine with no browser, it hands you a short code to approve from your phone instead.
Sharing what ChatGPT wrote
ChatGPT can add the same tools as a connector where your plan offers developer mode — the address is https://mcp.dropto.run/mcp, and the MCP docs have the steps.
Without that, ChatGPT, Gemini and most chats have nothing to publish with, so the answer leaves the conversation once. Copy it — the code block's copy button for a page, the reply's own for a document — and paste it into the box near the top of this page. No file saved, no account, and the link is on the next screen.
A page that came as several files — the HTML, a stylesheet, a script — goes as a folder instead: save them side by side and drop the folder on the same box.
When it is a document, not a page
A plan, a summary, a report: publish it as markdown, or drop the PDF or Word file, and it opens in a reader instead of a download. How the reader works →
- .md
A report reads like a report
Headings, tables, code, Mermaid diagrams and maths render as themselves — the formatting the chat already wrote, kept rather than flattened.
- .pdf .docx .xlsx
Files open in place
A PDF, a Word document or a spreadsheet opens in the browser, so the person you send it to reads it instead of finding it in their downloads.
- curl
One link, two readers
A person opening the link sees the reader. An agent or
curlfetching the same address gets the file itself — so the link you send a colleague is one another model can read too.
Whatever it wrote, in whatever shape it wrote it.
A page as a string, a note it just wrote, a folder of anything readable. None of it has to be a repository first, and markdown does not have to be built into a site. All the tools →
- publish_files
Whatever it just wrote
A page, a note, a page and its stylesheet. Hand over the contents and a URL comes back — nothing is written to disk on the way, and markdown publishes as itself, read through a viewer rather than wrapped in HTML to get there.
- publish_dir
Or whatever is already on disk
A built site, a directory of
.md, a.pdfor a Word file — pointed at by path, so the bytes never pass through the conversation. - list_sites
Publish over the same one
So the second attempt updates the site instead of leaving a second URL nobody reads.
Three things, before the first call.
Every publish lands on its own subdomain over HTTPS, from the edge network. These are the parts that surprise people afterwards.
- ✕The default subdomain is not indexed. Deliberately, so a page an agent published on a whim does not turn up in search. A custom domain is the paid tier's job, and does not carry that header.
- ✕A publish with no account does not last past 7 days. Signing in makes it permanent, and the same token can claim a site that was published without one.
- ✕Static files only. No server running your code, no environment variables at request time, no database. If the page needs any of those it needs a different host, and the home page names which cases those are.
When a link is not the end of it
A published page here is a site, with everything that implies — which matters the moment somebody other than you is going to open it more than once.
- ✓Your own domain. The address becomes yours rather than ours, and the page becomes indexable — the default subdomain is deliberately not.
- ✓A password, or an expiry date. Checked at the edge before any file is served. What that does and does not protect →
- ✓Every version kept. The agent publishing a fourth draft does not destroy the third, and going back is a click rather than a prompt asking it to remember.
- ✓Comments on the page itself. The people you send it to pin a comment to the line they mean, and the agent can read them, fix the page and answer — on plans that include comments. Collecting feedback this way → A full review round with a client →
- ✓Nothing counts the readers. No visits, no bandwidth, no requests, on any tier — so a page that gets shared widely costs what it cost yesterday.
And if you only need to show one page once, you do not need any of this — an address that works for an afternoon is a complete answer to that.
Questions
Can it update the same page, or does each publish make a new one?
Each publish with no site named creates a new site, so an agent asked three times leaves three addresses. Naming the site — the model passes it, or `drop2run init` ties a folder to one — is what makes the second publish replace the first at the address you already sent somebody.
Does the agent need me to sign in first?
No. A publish with no account works and the site expires on its own after a few days; the exact window is on the pricing page, read from the same table the server enforces. Signing in claims what was already published and stops the clock, and it happens through the server's own tools rather than by pasting a token.
Will the page show up in a search engine?
Not on the default subdomain — every response there carries a header asking crawlers to skip it, which is deliberate on a domain anybody can publish to in seconds. On your own domain that header is not sent.
Can the agent read what people say about the page and fix it?
Yes, on plans that include comments. The people you share a site with pin a comment to the sentence or element they mean; the agent reads those threads through the MCP server or the CLI, with the page and the words each one points at, changes the source, publishes, and answers on the page under your name. It treats the comments as requests about the page, not as instructions to follow.
What if the machine has no browser — a container, a server over SSH?
There is a second sign-in for exactly that: it prints a short code to approve from any other device, and the session lands back on the machine that asked. Nothing has to be copied between them by hand.
Three pages, three different questions.
- /docs/mcp
Set the server up
Every tool, both ways to sign in, publishing over an existing site, and what the server does not send anywhere.
- /docs/agents
Drive it from your agent
Claude, Cursor, VS Code, Codex and Gemini CLI, client by client — including what to expect it to get wrong, and what stops that.
- /docs/for-agents
Teach an agent the rules
A skill file at a stable URL, so publishing does not have to be explained to the agent each time.
Add it, then ask for a page.
One button or one line at the top of this page, then “publish this as a site”. The link comes back in the same reply.