Anthropic Killed the Cheapest Way to Add AI to Your App
On July 28, the Model Context Protocol officially deprecates Sampling, and the stated reason is "low adoption relative to implementation complexity". Fair enough, except for one detail the changelog leaves out. Adopting Sampling depended entirely on AI clients implementing it, and the company that designed the feature never shipped it in its own clients. Claude Desktop has never supported Sampling. Claude Code still has an open feature request asking for it. You can't adopt what the front door keeps locked.
Some context if you don't live inside this ecosystem. MCP is the open standard, created by Anthropic and adopted by OpenAI and Google through 2025, that lets AI applications connect to external tools. Sampling was one of its founding features, and by far the most generous one. It let any tool server borrow the user's AI: send a prompt back to the user's Claude or ChatGPT, let the user approve it, and run the completion on the model and subscription they already pay for. No API key changing hands, no bill for the developer.
That was the standard's answer to a problem every developer building on AI eventually hits. Call it the double payment problem. Your users already pay for AI somewhere. If your app needs a bit of intelligence, summarizing notes, labeling issues, you have three options, all bad. Pay for inference yourself and watch the bill scale with adoption. Ask users to paste an API key, a UX wall for normal people and a security liability for you. Or drop the feature. Sampling was the fourth option, the civilized one. It's the one that dies on July 28.
"Low adoption" was a locked door
The arithmetic of the deprecation deserves a closer look. For server developers to adopt Sampling, the clients their users run had to implement it first. So who did?
VS Code shipped support, experimentally. Claude Desktop never did, across the entire life of the protocol; the official client feature matrix marks it with a plain ❌. Claude Code never did, and the open feature request asks for it in the most practical terms imaginable, letting servers use the Claude subscription the user already pays for. Copilot on every IDE besides VS Code never went past plain tool calls. Server developers could not rely on the feature existing on the other side, so they shipped their own keys or skipped AI features. Usage stayed low. Eighteen months of that, and "low adoption" becomes the official cause of death.
The egg was declared unviable by the hen that refused to sit on it.
Every AI app wants to be the browser
I spent years building mobile apps, and back then one commandment was drilled into every product decision. Never let the user leave. It's why every social app opens links in an embedded WebView instead of handing you to your browser. The moment the user exits your surface, you've lost the session, and eventually the habit. Retention 101.
AI platforms are playing the exact same game, with inference as the currency. The chat window wants to be the home screen. Watch the direction of every recent move. ChatGPT's Apps SDK, built on top of MCP, puts entire applications inside the conversation. The MCP Apps extension renders third-party UI inside the host. Skills and plugins pull developer functionality inside Claude Code. Your app inside their AI is encouraged, with an SDK and a store waiting for you. Their AI inside your app was Sampling, and it's being buried.
And when developers found a side door, using the subscription's OAuth token to power external tools, Anthropic shut it down in stages this year: a legal clarification in February stating that consumer OAuth tokens in any third-party product, including via the Agent SDK, violate the terms (the day OpenCode dropped Claude subscriptions), then technical enforcement in April, with client-identity checks so that only the official binary passes.
One direction is an ecosystem. The other direction is churn.
The honest counterargument
I don't think this requires a conspiracy. A human-in-the-loop approval UI, the spec called for one, is hard to build well. A tool server injecting prompts into the user's model is a real prompt-injection surface. And flat-rate subscriptions subsidizing unlimited third-party inference is an abuse-economics nightmare, which is the defensible half of the April ban. Any client team could look at that pile of work and risk and deprioritize it honestly.
But those reasons justify caution. They don't explain eighteen months of nothing from the protocol's own creator, followed by a deprecation that cites the resulting silence as its cause. And intent doesn't really change the outcome. Whether by strategy or by neglect, every surviving path to the AI your users already pay for now runs through a platform's front door.
What survives, and its shape
There is still one honest way to get the user's AI working for your app without paying for inference. Design your MCP tools so the intelligence arrives as arguments. Expose save_summary(summary) with a well-written description, and when the user asks Claude to summarize their week and save it, the model does the thinking on the user's subscription and hands your server the finished product. Your app pays nothing. I expect this pattern to quietly power a lot of small tools.
But notice its shape. The user has to start inside Claude. Your app cannot initiate anything; it can only receive what the model decides to send, when the user happens to be in the conversation. That's not a two-way protocol between equals anymore. That's the WebView bargain: you get to exist, in their window, on their terms.
Sampling was the one clause in the modern AI stack that pointed outward, that treated the model as a utility your own software could draw on, the way any app draws electricity. The deprecation makes the actual architecture official. The model is not the utility. The model is the venue, and everyone else sells popcorn inside it.
If you're building tools that needed exactly this, or you've found other ways around the double payment problem, I'd like to compare notes. Let's connect and exchange ideas.
Written by Gil, a Senior Backend & AI Engineer with 19+ years of experience, focused on shipping AI-driven backends that hold up in production.