AI Engineering
Learn LLMs, RAG, agents, secure tools, and production AI through practical insurance examples.
What problem does MCP solve?
It standardizes how AI applications connect to and discover external tools and context. Multiple approved hosts can reuse the same server-side capability contracts rather than each building a custom integration.
MCP versus function calling?
Function calling lets a model propose a structured operation. MCP provides an integration protocol for discovering/invoking capabilities. The host can turn a model proposal into an MCP request that calls an existing REST service underneath.
How do you secure an MCP server?
I authenticate requests, verify token properties, authorize each tool and object, restrict downstream credentials/egress, validate schemas/results, review tool definitions, and audit actions. Connection permission does not equal customer-record permission.
A new discovered tool requests secret access. What happens?
The host should not automatically expose or execute it. Tool descriptions are untrusted unless reviewed; allowlists and execution policy block capabilities outside the task. I investigate the server/definition change and preserve an audit record.
Difficulty is an editorial estimate. Completion means you chose to mark a question as understood.