My site reads fine to humans. To an AI agent it's a wall of HTML it scrapes badly — or skips. So I gave it a front door: one file, one command, and any Claude knows everything on it.
I asked a fresh Claude "what has Jacob built?" and pointed it at tellingtechnology.skogstrom.se. It fetched the page, choked on the aurora-and-animation markup, and gave me a vague half-answer. The content is all there — it's just buried in HTML written for eyes, not for agents. The web has a fix for this now: llms.txt, a plain-text map of a site meant for language models.
I don't hand-write the map — the site builds it. A small script walks the showcase pages and emits two things on every deploy: llms.txt (a readable index of episodes, hooks and tools) and index.json (the same data, structured). Because it's generated, it can never drift from the real pages. The site that builds itself at night now also describes itself to machines.
Two ways in, same index behind both. npx tellingtech search "globe" answers from your terminal — no install. And a thin MCP server wraps the same JSON so any Claude that adds it can search episodes, fetch a script, list tools used as first-class tools. Built on Anthropic's Model Context Protocol — the same standard everything else in this series plugs into.
The MCP server only ever reads the public index — no write tools, no secrets, no account access. An agent can learn what I've built; it can't touch the site. The whole surface is the same index.json you can open in a browser.
The build in four frames — the CLI, the registered server, the split-screen demo, and the before/after that makes the whole point. Tap any image to enlarge it and read the exact prompt that drew it.
New Claude session. Add the MCP server. Ask the question that failed in beat 01 — and this time it lists every episode, the hook, and the tools, straight from the live site. Same question, two worlds apart.
Three minutes: the failed scrape, the generated llms.txt, the MCP server going live, and the one-shot answer.
The generator, the CLI, and the MCP server — point them at any static site and it starts explaining itself to agents. Everything in this episode is free and open — clone it, run it, make it yours.
npx github:tellingtechnology/tellingtech-mcp search "globe"
No GitHub? Comment FRONTDOOR on the post and the bot DMs you the link.
A single terminal strip that shows context left and tokens burned, live, while I work. The cockpit for an agent that never sleeps.
