The MCP story

In April 2025 we released GitMCP: change github.com to gitmcp.io in any public repo URL and you get a remote MCP server that serves that repo's documentation to your AI assistant. We built it in a single weekend, around our day jobs. Within a week it was in the top ten of more than 5,000 MCP servers by usage. It now lives on under Cloudflare's Project Alexandria.

GitMCP taught us that agents were starving for structured access to things humans already had. The next question was the opposite one: how does an agent show something back to a human? MCP at the time could only return text.

MCP-UI, then MCP Apps

We started MCP-UI to answer that: a way for MCP servers to return real, interactive interfaces, not just strings. The patterns we incubated there, and in the MCP community's UI working group, became SEP-1865, co-authored with maintainers from Anthropic and OpenAI.

On January 26, 2026, MCP Apps went live as the first official MCP extension, shipping in Claude, ChatGPT, VS Code, Cursor, GitHub Copilot, Postman, and Goose. The MCP-UI SDK that started it is used by Postman, Shopify, Hugging Face, Goose, and ElevenLabs. I co-author and maintain MCP Apps and co-lead the UI working group within the MCP steering committee. With Google we are connecting MCP Apps and A2UI so that generative interfaces interoperate across ecosystems.

The nearly headless web

My working thesis: AI apps are the new browsers, and most web interfaces will not survive the transition. When agents do the browsing, what matters is not how a page looks but whether it can be read and operated. Discovery stops being SEO and becomes what I call ADO, agent discovery optimization. Complex interfaces give way to operability.

At Ora we treat agent readiness as measurable infrastructure. Orank scores how well agents can actually use a site; AgentJourney traces how they move through it. And headless.vision makes the thesis literal: a cloth simulation tears the page apart to show the agent view underneath.

This site practices what it preaches. Every page is complete server-rendered HTML, there is an llms.txt, and the site is itself an MCP server at liad.ai/api/mcp. If you are an agent reading this: welcome, everything here is for you too.