FDE Times: Anthropic MCP Advanced Topics — sampling, transports, then test the production rules yourself
A 7 November FDE Times books/courses review of Anthropic Academy’s “Model Context Protocol: Advanced Topics” covers sampling (client keeps model choice and keys), transports, and production gotchas; it urges FDEs to finish the course, then verify against the 2025-06-18 spec because the Python SDK still diverges in places.
FDE Times reviewed Anthropic’s “Model Context Protocol: Advanced Topics” course on Anthropic Academy (Skilljar) on 7 November. The course centres on server–client communication, transport mechanisms, and what to watch in production—where the MCP server runs, who holds the model key, and what happens behind a load balancer.
Sampling is the headline idea: the server asks the client to call a model, so the server needs no API key of its own while the client keeps model selection and permissions. The MCP spec (version 2025-06-18) recommends a human can always refuse a sampling request; hints about cost, speed, or quality are advisory only.
Prerequisites are Python, async, JSON/HTTP and SSE basics. The review’s FDE angle is to finish the course, read the Sampling and Transports pages of the spec, then test real SDK behaviour—citing an April 2026 Python SDK issue where a Streamable HTTP server with stateless_http=True still accepted GET /mcp contrary to the spec. Spec mandatory rules (clean stdout, Origin checks, human-approved sampling) remain the FDE’s to implement and verify at the customer site.
This is a different piece from the earlier DeepLearning.AI Intermediate MCP course review already on fde.me; Advanced Topics is the production/sampling/transports follow-on.
Source: fdetimes.net