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 block/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