Pageree

An AI assistant can write a good landing page in minutes. Then the person spends an hour on the rest: hosting, a form that actually sends, analytics, and something to carry the lead to an inbox. None of those tools was built for an agent to operate, so the person does it by hand, every time.

Pageree is built for that last mile. It is a hosted MCP server. Connect it once, with OAuth, from Claude, ChatGPT, Cursor, Codex, Claude Code, VS Code or Gemini CLI, and the assistant gets the tools it was missing: preview a draft privately, publish to a live URL or a custom domain, add a form that is validated and spam-protected, deliver every lead to email, a webhook or a CRM, read section-level analytics, and roll back to any earlier version. The person never leaves the chat.

Building a product with no real UI of its own changed more than expected. Six things worth sharing.

1. The user is the model, and the model reads the docs.

In a web app, onboarding is screens. In an MCP server, onboarding is the text of the tool descriptions. If "publish" does not say a draft must be previewed first, the agent publishes an unreviewed page. If nothing says to keep the page id, the agent creates a second page instead of a new version of the first. That single bug cost a week; the fix was one sentence.

The answer was a QA harness: a fresh agent with no memory, the tool descriptions, and the task "get a landing page live for this business". It publishes a real page to a test account and the transcript gets read. The success rate of a naive agent, from descriptions alone, is now the metric releases are judged by.

2. Skills are the product. Tools are the plumbing.

Anyone can wrap a hosting API in MCP. What makes the pages good is the instructions the server hands the agent on demand: how to structure a page that converts, how to write the copy, which fields belong in a form, how to read the analytics and what to change. It felt like documentation rather than software. It is the part users notice.

3. Chat users have no filesystem.

A developer in Claude Code has a project folder, so state can live in a file. A ChatGPT user has a conversation. Everything the agent must remember across sessions, such as which page it was working on, has to come back from the server in a form the model reliably keeps. Structured tool output stopped being optional.

4. The directories have opinions about your features.

To be found inside ChatGPT or Claude, a connector goes through their directories, and the rules reach into the product. One does not accept connectors that generate images with AI models; Pageree has an image tool because Claude cannot make images and ChatGPT cannot pass its own into a tool. Another forbids selling subscriptions in the chat, including a link to checkout, so the trial has to get a person to a published page with a working form before they ever see a price. Those are product decisions a web app would never have forced.

5. The moat is not the loop. It is the data in the loop.

Every layer has a cheaper incumbent: Netlify, Tally, Plausible, Zapier. Wix, Webflow and Hostinger shipped ChatGPT apps this year, and ChatGPT can now host a page itself. "Build and publish" no longer sets anything apart. What none of them do is give the agent section-level behaviour, which section people leave on, how far they scroll, what they click, as text it can reason about, and then let it edit one section, republish and measure the difference. Whether people want an agent iterating on a live page from its own analytics is the one assumption still unproven. Everything else is built.

6. Pricing has to fit in a sentence a model can repeat.

The user never visits a pricing page; they connect from a chat. So: 30-day trial, then $20 a month for 5 pages or $40 for 30, unlimited form submissions. Pages are the cap because it is the number everyone understands. It may not be the right cap.

If Pageree were started again: the QA harness on day one, the skills before the tools, and the directory policies read before choosing features.

Pageree launched on Product Hunt this week, built by a solo founder. Two questions for people here who have built for an agent as the user rather than a person: what surprised you, and is a page cap the right way to price something like this?