Explainer

What MCP is, and what it actually lets your AI assistant do

MCP is the plug that lets an AI assistant use tools outside its own chat window. What it is in plain terms, what it changes about your workflow, and how publishing a page through it works.

You keep seeing "MCP" next to the name of every AI tool — connectors, integrations, servers, tool calls. The acronym gets explained in one of two unhelpful ways: as a protocol spec you have no reason to read, or as a marketing word that means "our thing works with your thing".

Here is the useful version, and what it changes about the boring part of your day: getting work out of the chat window and in front of another person.

MCP in one paragraph

MCP — the Model Context Protocol — is a standard way for an AI assistant to use tools that live outside itself. Anthropic published it in late 2024 and every major assistant has since adopted it. Before MCP, if an assistant wanted to read your calendar, query your database, or publish a file, someone had to build a bespoke integration for that one assistant. MCP replaces the pile of one-off integrations with a single interface: a tool describes what it can do, the assistant reads that description, and it calls the tool when it's useful.

The comparison people reach for is USB. Before it, every device had its own connector. After it, one port fits everything. MCP is that for AI tools — the assistant is the port, and each MCP server is something you plug in.

Two words are worth pinning down, because the docs use them constantly:

  • An MCP server is the tool being offered. Pagepost is one. It isn't a machine you run; for remote servers it's just a URL.
  • An MCP client is the assistant doing the calling — Claude, ChatGPT, Cursor, and so on. Some call the same thing a "connector".

There are two flavours of server, and the difference is the one thing worth knowing before you try to set anything up:

  • Local (stdio) servers run as a program on your own machine. Good for touching your filesystem or a database on localhost.
  • Remote (HTTP) servers are a URL somewhere on the internet, with a sign-in. Nothing to install, works from a phone, same for everyone on the team.

Pagepost is a remote server. That's why setting it up is pasting a URL and signing in, not installing anything.

What changes in practice

The honest answer is: one specific kind of friction disappears — the copy-paste-download shuffle at the end of a task.

Without a connector, asking an assistant for a report goes like this. It writes a good HTML dashboard. You copy the code, paste it into an editor, save it as dashboard.html, open it to check it survived, then upload it somewhere that gives you a URL, then paste that URL into Slack. Six steps, all of them yours, every single time. And when the numbers change tomorrow you do all six again, and the person you sent it to is still looking at yesterday's link.

With a connector, you say "publish that" and get a URL back. The assistant does the file handling because the file never needed to be your problem.

That's the whole value, and it's worth being clear about the scale of it: MCP doesn't make the assistant smarter. It removes the manual steps between "the assistant made a thing" and "someone else can see the thing".

What Pagepost offers over MCP

Pagepost is a remote MCP server that turns whatever your assistant wrote into a real, shareable URL. It exposes four tools, and the assistant picks between them on its own:

Tool What it does
publish_file Takes an HTML or Markdown document and returns a live URL
update_file Rewrites a page you already published, keeping the same URL
search Lists the pages you've published, so the model can find one by name
fetch Returns a page's real source so the model can revise it

The first one is the one you'll use daily. The other three are what make it more than a one-way upload:

Interactive HTML keeps working. An HTML page is served as you wrote it, scripts included, so charts, filters and tabs still do what they did in the preview. Markdown gets rendered to a clean styled page with a table of contents instead of arriving as a wall of # symbols.

The link is editable, and stays the same link. This is the part people don't expect. Because fetch can hand the model the real source of a page you published last week, you can say "pull up the Q3 dashboard and add the September numbers" — it fetches, revises, calls update_file, and the URL you already sent to twelve people now shows the new version. No second link, no "ignore the one I sent earlier".

One library, whatever published into it. Pages published from Claude sit in the same dashboard as pages your colleague dropped in Slack and files you uploaded from the web. Rename, revoke, set an expiry, check the view count, or fix a typo directly on the live page.

Scoped to your account. The connector signs you in once through OAuth, and every call after that runs as you. A model can only search and fetch pages you own — asking for someone else's page gets the same answer as asking for one that doesn't exist.

How to use it, end to end

The setup is a URL and a sign-in. Once, per client.

  1. Add the server. In your assistant's connector or MCP settings, add a remote server with the URL https://pagepost.app/api/mcp. In Claude's custom-connector dialog, https://pagepost.app works too.
  2. Sign in. A browser window opens; sign in with Google or Slack. There's no API key to generate or paste — Pagepost issues tokens through OAuth only.
  3. Ask for something. Prompt as you normally would: an interactive dashboard, a client one-pager, a Markdown status report.
  4. Say "publish it to Pagepost". The assistant calls the connector and puts the live URL straight into the chat. Paste it anywhere a link works.
  5. Revise later in the same breath. "Load the pricing page from Pagepost and fix the VAT line" — same URL, updated content.

The exact settings screen differs per client, which is the only thing that does. There are step-by-step pages for Claude, ChatGPT, Gemini and DeepSeek, and a connector reference with the endpoint, transport and tool list for everything else — Cursor, VS Code with Copilot, Cline, Zed, LibreChat and friends.

What MCP won't fix

Worth knowing before you go looking for a setting that isn't there.

  • Some assistants can't add remote servers at all. DeepSeek's own chat app has no MCP settings; the workaround is to run a DeepSeek model inside a client that does.
  • A client that only takes a pasted API key can't connect. Pagepost authenticates through OAuth, so there's no long-lived key to paste. That rules out a handful of otherwise capable clients.
  • The model still decides when to call a tool. It usually gets this right, and naming the tool ("publish it to Pagepost") removes the ambiguity when it doesn't.
  • A connector is not a substitute for judgement about what you publish. A page URL is unguessable, not access-controlled. If a document genuinely shouldn't be readable by anyone holding a link, restrict it to your Slack workspace, set an expiry, or don't publish it at all.

The short version

MCP is a standard plug that lets an AI assistant use outside tools, so the assistant can finish a job instead of handing you a file to deal with. Pagepost plugs in as a remote MCP server: your assistant publishes what it wrote to a real URL, can pull that page back to revise it later without the link changing, and everything lands in one dashboard you control.

If you want the technical specifics — transport, OAuth details, tool schemas — they're on the MCP connector page. If you'd rather skip connectors entirely, dropping the file on pagepost.app gets you the same link with no setup.

Send your next page as a link

Drop an HTML or Markdown file in and get a URL that renders it for anyone you send it to. Free, and it works from Claude, ChatGPT, Gemini and Slack too.

Try it free

More guides

Send feedback

How are you feeling about Pagepost?

Choose which optional categories you allow. Strictly-necessary storage is always on — without it, the site cannot keep you signed in.

Strictly necessary

Required for sign-in, OAuth security, and remembering this cookie choice. You can't turn these off.

Analytics — Google Analytics

Aggregate usage metrics — page views, traffic sources, broad engagement patterns. Runs in Consent Mode v2: when off, no cookies are set and Google only receives anonymous, cookieless aggregate signals.

Session replay — Microsoft Clarity

Anonymised replays of how users navigate the site, used to find UX issues. The Clarity script is not loaded at all when this is off.

See our privacy policy for full details, including international transfers and your rights.
What MCP is, and what it actually lets your AI assistant do