MCP Best Practices
summary
Defines MCP as a local MCP server that the model calls via JSON over RPC, implemented here with a Python script and a “fast MCP” style library that registers callable functions and logs requests; demo shows translation and the LLM stripping extraneous characters and working end-to-end. Covers RAG for embedding subsets, agentic actions (APIs that book tickets), and risks: untrusted servers, leaked API keys, need to rotate creds, audit logs and validate returned data. Also mentions small tuned models around 30 billion parameters.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:24Why MCP for LLMs
- 3:09Extending LLM capabilities
- 5:41RAG with LLMs and embeddings
- 6:23Model Context Protocol explained
- 9:35MCP beyond data retrieval
- 9:59Developing MCP server
- 13:31Live demo: MCP translation
- 16:46Comments in MCP functions
- 17:09MCP security risks
- 19:41Validating MCP data
- 20:01Small models and tuning
related talks
35mMCP, or not MCP
22mSAFE-MCP: A Security Framework for AI+MCP (Model Context Protocol)
58mMCP with .NET: securely exposing your data to LLMs
26mGovernance and Security of APIs and MCPs
15mRemote MCPs: What we learned from shipping
25mYour Insecure MCP Server Won't Survive Production
52mMCP, Context Model… Perfection?
50mMCP at Scale: Governing AI Agent APIs When Autonomous Systems Are Your Consume…