Three things I’ve been watching, and what the public record actually says about each.
1. The assistant is becoming a distribution channel
Companies are starting to treat Claude and ChatGPT the way they once treated the App Store: a place where customers arrive, not just a place where their product gets described.
Canva is the clearest case I can find. It shipped an MCP server in July 2025 connecting to Claude, extended brand-kit support to both Claude and ChatGPT in February 2026, and in June 2026 full design creation landed inside ChatGPT. Canva says those integrations brought in millions of new users without denting engagement on its own site. For a company at 265 million monthly actives and $3.5 billion in annual revenue, an acquisition channel that size is not a side project.
Canva’s position is unusual, though, and the reason matters. Canva sells a capability. When an agent calls it, the agent does the selling and Canva does the work. If what you sell is content, that same agent is a substitute rather than a front door. Published crawl-to-referral ratios differ sharply by source and by month, but every version of the number points the same way: AI platforms fetch hundreds to thousands of pages for each referral they send back.
The first observation holds, then, with a condition attached. The assistant is an entry point for products that do something, and an extraction layer for products that only say something.
2. The products are shifting, and the shift has a name
Netlify’s Mathias Biilmann called it Agent Experience, or AX, and framed it as a survival question. Companies should “start consciously designing the AX of their products, or risk being replaced by tools that empower their customers to harness the exponential power of seamlessly collaborating with agents.”
The traffic numbers back the premise. Cloudflare reported this year that more than half of internet traffic is no longer human, and that daily requests from AI agents on its network grew by more than 1,700% over the past year. Your second-largest user population does not render CSS.
What I underestimated is how fast “be agent-friendly” stopped being a design philosophy and became a procurement decision. A merchant who wants to be bought by an agent now has to pick rails: OpenAI and Stripe’s Agentic Commerce Protocol, Google and Shopify’s Universal Commerce Protocol (launched January 2026 with Etsy, Target, Walmart and Wayfair behind it), Google’s AP2 for proving an agent was authorised to pay, Coinbase’s x402, with MCP underneath for catalog and cart. Most production deployments run two or three together, because none of them covers the whole path from discovery to fulfilment.
And the door itself is missing. Nothing crawls for MCP servers. Every place your server can appear is a submission you file by hand. The July 2026 spec added a discovery call, but a client only invokes it once it already has your URL, so it is not a directory. Server cards became an optional extension in October 2026, and no client is documented as using them to find servers yet.
That is the part I would flag to anyone planning this work. The agent-native channel has the economics of a platform and the plumbing of a 1998 web ring.
3. “The best UI is no UI” turned out backwards
Golden Krishna’s 2015 book argued that the best interface is no interface, and the line has had a second life as a description of agents. The industry went the other way.
MCP Apps became the first official MCP extension in January 2026, co-authored by people from Anthropic, OpenAI and the MCP-UI project. A tool declares a ui:// resource holding an HTML page. The host renders that page in a sandboxed iframe under a deny-by-default content policy, and the page talks back over JSON-RPC. Claude, ChatGPT, VS Code’s Copilot agent and Goose all render them. The destination turned out to be interface delivered through the conversation.
The distinction worth holding onto is who writes it. In MCP Apps the developer authors the interface ahead of time and the agent chooses it, which is closer to responsive design than to generation. Genuinely generated UI exists too. FastMCP’s generative provider has the model write component code at runtime, streams it into a browser sandbox as it is written, then revalidates the finished code server-side. It works, inside a fixed component vocabulary and a sandbox with no third-party libraries.
So “an intelligent UI on the fly” is two different products. One is shipping and standardised. The other is real but bounded, and the bounds are where the engineering actually is.
What I would watch
The decision in front of most teams is not whether to support agents. It is narrower and more awkward: which assistant’s distribution you are willing to depend on, and what happens when that platform ships your feature natively or changes the access terms. Canva is the case to study in both directions. The reason the integration works is that the agent does the selling and Canva does the work – which is also the reason the platform, not Canva, now owns that customer relationship. Millions of users arrived through a channel Canva does not control and cannot price.
