Anthropic Ends MCP Sampling

melanke · reddit · 2026-07-15

The author analyzes Anthropic's plan to deprecate MCP Sampling on July 28, questioning their cited "low adoption rate."

What Sampling Originally Did

Allowed servers to leverage models within the host client for generation.
Servers didn't need their own API key, theoretically letting developers plug AI features "directly into a user's existing subscription."

Why the Author Thinks It Never Truly Took Off

This capability required implementation on both the client and server sides.
However, crucial client support was extremely limited:
VS Code: Only experimental support
Claude Desktop: None
Claude Code: None
Copilot (outside VS Code): No real support either
As a result, server-side developers couldn't rely on it as a stable capability, and the ecosystem naturally cooled down.

The Author's Assessment

This isn't necessarily a conspiracy; it's more likely that human-in-the-loop approval UIs are hard to build, compounded by the fact that it inherently presents a prompt injection risk surface.
But these reasons explain "caution," not why the protocol and flagship clients saw almost zero action for so long.

Patterns Still Worth Borrowing

If the schema is designed to let the model think first and then hand the conclusion over to the server, some scenarios remain viable.
It's just that this approach is better suited for initiating from within the host chat interface, rather than letting your application completely control the flow.

More from coding & agent