MCP App Sampling

[alexhancock]

Pre-submission Checklist

• I have verified this would not be more appropriate as a feature request in a specific repository
• I have searched existing discussions to avoid duplicates

Your Idea

I have an implementation of sampling from app code working in goose aaif-goose/goose#7039

My motivation is to make MCP apps themselves able to be AI-powered, leveraging the host's model connection

A detailed proposal is available here https://gist.github.com/alexhancock/2bb8b62500f33ccb29a0edb4aa7c5342, and to summarize:

• MCP Apps (Views running in iframes) can currently call tools and read resources, but have no way to request LLM completions from the host
• I propose adding a sampling/createMessage message that lets Views request LLM completions, mirroring the existing MCP sampling capability that servers already have
• The host advertises sampling support during initialization; Views check for the capability before making requests
• Views send conversation messages, an optional system prompt, and an optional max token count — the host routes it to its LLM provider and returns the result
• The host stays fully in control: it chooses the model/provider, enforces limits, etc
• Enables new app patterns like embedded chat UIs, AI-powered form completion, content generation, interactive tutoring, and data analysis assistants
• A working reference implementation exists in Goose: block/goose#7039

Scope

• Protocol Specification
• SDK Features
• Documentation
• Developer Experience
• Other

[hydrosquall]

I'm interested in offering this capability, but I noticed the latest MCP release candidate is deprecating sampling, though I didn't fully follow the reasoning for it.

https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/#roots-sampling-and-logging-are-deprecated

As a workaround, I've added a custom server tool to my MCP server which allows the UI to access an LLM provider, but it's chosen by the MCP server rather than something that the user could customize in their agent.