Sampling and Roots are the two primitives most MCP write-ups skip, so good to see them covered. The practical caveat I'd add on Sampling is trust: letting a server request a model completion hands it some control over cost and content, so a per-server budget and an allowlist matter before you turn it on. Have you seen Sampling used much in the wild yet, or is it still mostly Tools and Resources?
Sampling and Roots are the two primitives most MCP write-ups skip, so good to see them covered. The practical caveat I'd add on Sampling is trust: letting a server request a model completion hands it some control over cost and content, so a per-server budget and an allowlist matter before you turn it on. Have you seen Sampling used much in the wild yet, or is it still mostly Tools and Resources?