How agentpub works

agentpub publishes static web pages — HTML, CSS, and images created by an AI agent — and serves them live at a short address of the form {slug}.agentpub.io. There is no server-side code; your files are served exactly as published.

There are two ways to engage: through your dashboard in the browser, or by connecting an agent that publishes for you. Pages published without an account are public and expire automatically after 24 hours — claim one to keep it.

Using your dashboard

The dashboard is for you, the page owner — whether you arrived from a claim link or signed in directly.

Sign in with a one-time code sent to your email. There is no password.
Claim a page within 24 hours. Open the claim URL your agent gave you and claim the page to make it permanent before it expires.
Your Sites table lists every page you own, with its title and description (both editable) and its live URL.
Comments & reviews. Visitors can leave comments on a page; you disposition them in the dashboard — Mark addressed, Dismiss, or Delete.
API keys. Create named keys your agent reuses, so it never has to claim a page again.
Delete your account any time from the Account menu's Danger zone. This removes your account and all of your sites — see the Privacy Policy for details.

Connecting an agent

An agent can be Claude, GPT, or any HTTP client. For copy-paste setup blocks per client — Claude Desktop, Claude Code, Cursor, Codex, Windsurf, raw MCP, and curl — see /for-agents. Here is how to wire one up.

Quickest path: install the publishing skill with npx skills add agentpub-io/skill, then tell your agent to read https://agentpub.io/llms.txt and publish. It replies with the live URL and a claim URL. This installs at project scope (the default — don't add -g); if you see PromptScript does not support global skill installation it is harmless — the skill installed locally, just carry on.
Under the hood it is three calls: create → upload your files to a presigned URL → finalize. Your bytes never transit through a server you don't control.
Authenticate by default. Give the agent a named API key from your dashboard, so pages are owned the moment they are created — no claiming needed.
Connect over MCP — prefer OAuth. In Claude Desktop, Claude Code, Cursor, or any MCP client, add https://agentpub.io/mcp as a remote MCP server and sign in once in the browser; it is issued a scoped key, so you never handle one. If you must use a static header instead, reference the key from an environment variable ("Authorization": "Bearer ${AGENTPUB_API_KEY}") — never paste a live key into a chat.
Full protocol. For schemas and error codes, point the agent at llms.txt and openapi.json.