LLMs, Meet Your Tools: The Magic of Model Context Protocol
summary
Model Context Protocol (MCP) lets an LLM request and run real tools: MCP server executes tools via JSON-RPC, advertises capabilities, and returns resources, prompts, roots and sampling decisions. Demos use Kubernetes, Playwright, filesystem, fetch and bash; an agent found and fixed two Docker issues (missing pandas and a bad base image), built the image and ran the Python script successfully. Entropic and OpenAI adopt MCP and report detection of prompt injections but Entropic reports 88% prompt rejection, so tool poisoning, shadowing and local file access remain real risks.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:08LLMs need grounding
- 3:55Model Context Protocol
- 5:24MCP server/client
- 6:09MCP features
- 12:15Popular MCP servers
- 14:00MCP tools in IDE
- 16:07Custom MCP servers
- 17:34Live demo: MCP client
- 20:26MCP security risks
- 24:21Prompt injection attacks
- 28:37MCP exploitation example
- 30:05Protecting from MCP risks
- 33:14Q&A
related talks
43mLLMs, Meet Your Tools: The Magic of Model Context Protocol
32mMCP for dummies: Feed more context to LLM
46mMCP for LLM Agents, Infrastructure APIs & Data Services
50mBeyond local tools: Dive into the Model Context Protocol (MCP)
44mBeyond the Prompt: How MCP Connects Language Models to Real Tools
33mThe Model Context Protocol: Connecting AI to Everything
44mSupercharging DevOps with MCP (Without Opening a Security Hole)
26mBuilding AI-Enabled Spring Applications with Model Context Protocol (MCP)