My directory talks to agents

{
  "theme": "Cybersecurity",
  "feeds": 33,
  "ai_optout_excluded": true,
  "import_fornews": "https://ok-ia.ch/fornews/annuaire/ia-theme-cybersecurite.fornewssource",
  "page": "https://ok-ia.ch/fornews/annuaire.html#cybersecurite"
}

A real session, 8 October. The prompt is shortened; the JSON is what the page actually returned. No scraping, no screen control — the page answered.

cat the-problem.md

Until now, an agent that wanted something from your site had one method: load the page, read the HTML, guess what the buttons do, and click. It sort of works. It breaks the day you rename a CSS class. And above all, it asks a machine to do archaeology where answering would do.

WebMCP turns that around: the page declares its tools, and the agent calls them. It's a W3C draft from a community group, already implemented in the ChatGPT desktop app and in origin trial in Chrome and Edge. Ten lines are enough:

await document.modelContext.registerTool({
  name: "search_feeds",
  description: "Search 1,716 verified news feeds by keyword, language and theme.",
  annotations: { readOnlyHint: true },
  inputSchema: { type: "object", additionalProperties: false, /* … */ },
  async execute({ query, language, theme }) {
    // … answer from what the page already holds in memory
  }
});

The best part: it cost nothing. My directory already keeps 1,716 sources in memory to build its cards — the tools only declare what was already there. No extra request, not one byte for an ordinary visitor.

diff expected.txt actual.txt

Then I tried it. And nothing happened. I asked ChatGPT which tools it could see on the page, and it proudly answered:

"Chercher une source", "Toutes les langues", "Français · 432",
"Masquer les 743 fermés aux plates-formes d'IA", "Ajouter",
"Copier le flux", "Zapping"…

Those are my buttons. It was reading the DOM, exactly like before. A little later it even asked permission to drive my screen — the very thing WebMCP exists to avoid. My tools were there, alive in the browser, and completely invisible.

cat the-four-details.md

The answer was in OpenAI's documentation, which I had made the mistake of not reading before writing code. Four details. None of them is cosmetic: these are the signs by which an agent recognises a tool it may call without asking. Here they are — they cost me an evening, have them for free.

And a fifth, which was my fault rather than the documentation's: I had put the tools only on the page that holds the data. Agents navigate. It landed on the directory, followed a theme link, and lost everything on the way. All 38 theme pages carry them now, with the data fetched on demand.

cat it-works.md

After the fixes, the same question. This time it names search_feeds, list_themes, theme_bundle. And when I ask it to call the third one:

{
  "theme": "Cybersecurity",
  "feeds": 33,
  "ai_optout_excluded": true,
  "import_fornews": "https://ok-ia.ch/fornews/annuaire/ia-theme-cybersecurite.fornewssource"
}

That URL exists nowhere: not in the HTML, not even in my JavaScript file. It is assembled at call time. You cannot get it by reading the page — which is the proof the tool actually ran. I grinned at my screen like an idiot.

cat what-really-matters.md

Here is the detail I'm proudest of, and it isn't technical. Of the 1,716 publishers in that directory, 743 refuse AI crawlers in their robots.txt. I know because I've been collecting their robots.txt for months.

These tools are called by an AI. Staying quiet about that refusal would be a way of working around it.

So every answer says so, source by source, and one flag leaves them out — the bundle it returns then skips those publishers entirely. I don't filter by default: the refusal targets training and crawling, not a human's RSS reader. But saying it is the least one can do.

One last principle I care about: all my tools are read-only. The site answers, it never acts. It hands back the address of an import file; the app, on your machine, decides whether to add the source. The public site gets no write powers at all — and never needed any.

cat try-it.md

You'll need the ChatGPT desktop app, updated, its built-in browser, and a model that supports this — the docs name GPT-5.6 Sol or GPT-6 Sol, and say WebMCP is disabled on Luna. Open the directory and ask something like:

Find me French cybersecurity feeds I can add to my RSS reader,
skipping publishers that refuse AI crawlers.

In Chrome, enable about:flags#enable-webmcp-testing first. And to check in five seconds that the API is there at all, open the console: typeof document.modelContext.

play openai-walkthrough.mp4

OpenAI's own walkthrough, five minutes, and the clearest explanation of the idea I've found. Nothing is loaded from YouTube until you press play — the thumbnail is served from here, and the player only appears when you ask for it.

“Build custom site annotations with ChatGPT” — OpenAI

Further reading: the WebMCP specification · OpenAI's documentation · their video.