Dev.to AI 🤖 Ai 👁 0 📖 4 min read

Three tool calls: what happens when you ask Claude or ChatGPT to publish a press release

You type one line into Claude or ChatGPT: "Write a press release about our new booking page and publish it." A minute later you have a live news page with its own link. This post shows what happens in between. It is a w

You type one line into Claude or ChatGPT: "Write a press release about our new booking page and publish it." A minute later you have a live news page with its own link.

This post shows what happens in between. It is a walk through the LivePublicator MCP server: which tools the assistant calls, in what order, and why we built it so that a normal release takes three calls.

The setup, once

LivePublicator is a remote MCP server. The address is https://livepublicator.com/api/mcp (streamable HTTP, OAuth 2.1). You add it as a custom connector in Claude or as an app in ChatGPT, sign in once, and the assistant gets nine tools.

You don't need to sign in to try it. One tool, preview_press_release, works without an account and makes a private preview page that expires after 7 days. It publishes nothing.

The three calls

1. get_account

Before the assistant drafts anything, it calls get_account once. The answer holds the company, the saved press contact, the saved spokesperson, the credits left, and whether releases go live after the automatic checks or wait for a person.

This one call removes most of the back and forth. The assistant no longer asks "who is your press contact?" every time. If a contact is saved, create_release fills it in.

2. create_release

The assistant writes the whole release itself, from the chat or from a page you point it to:

  • a headline (10 to 200 characters, with the company name in it)
  • a one-sentence summary
  • a body of 250 to 1,000 words
  • 1 to 8 key facts
  • up to 6 questions and answers
  • a quote, with the person's name and role

Then it sends everything in one call. Here is a short example (the company is made up):

{
  "headline": "Acme Bakery opens online booking for custom cakes",
  "summary": "Acme Bakery now takes custom cake orders through a booking page on its website.",
  "body": "Acme Bakery today opened a booking page for custom cakes...",
  "key_facts": ["Bookings open 14 days ahead", "Pickup at the Main Street shop"],
  "faq": [{ "question": "How far ahead can I book?", "answer": "Up to 14 days." }],
  "quote": "Our customers asked for this for a long time.",
  "quote_name": "Dana Levi",
  "quote_title": "Owner"
}

The server checks length, structure, links, image size and content, and it returns every problem at once. That part matters for agents. If the server stopped at the first error, the assistant would fix one field, resend, hit the next error, and burn five calls on one draft. With every problem in one answer, it fixes them all and resends once.

If nothing is wrong, the answer is a private preview link and an edit link. Nothing is published yet, and no credit is used.

3. publish_release

The assistant shows you the preview and the key facts it wrote. Only after you clearly say "publish" does it call publish_release, with confirm: true.

That flag is on purpose. The tool is marked destructiveHint: true in its annotations, and its description says to call it only after the user has seen the preview. The assistant can't publish your news just because it decided to.

The answer comes back within seconds. It has the live link, or it says the release waits for a person to check it, with the reason and the expected time (up to 24 hours). That's why the assistant doesn't need a fourth call to check the status.

What you get

The live page has the release, key facts, the questions and answers, a press contact, and structured data, on your company page in our newsroom. Search engines are told the same day.

If you spot a typo after it goes live, fix_live_release can fix the text within 24 hours. The link stays the same, and the earlier version is kept.

What we learned writing the tools

  • Tell the assistant what not to call. list_releases and get_release_status both say "only when the user asks". That line keeps the assistant from calling them before every draft just in case.
  • Return all problems at once. One round trip instead of five.
  • Make the assistant draft, not the user. The server instructions say "Do not ask the user to write these." The user gives the news, and the assistant writes the rest.
  • Put the human yes in the schema. confirm: true on the only tool that publishes.

Try it

The full tool list, auth details and examples are in the docs: livepublicator.com/docs/mcp. Your first release is $9 (one per company), and drafts and previews are free.

This post was written with help from AI and checked by our team.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.