Prompting LLMs to use Tools and Resources with MCP
summary
Shows that small open models can be coaxed to call MCP tools and external APIs by embedding concise JSON/YAML schemas and Llama-style instruction templates, and that a homemade MCP test server plus a testing app actually return parseable JSON tool calls. Finds YAML often more reliable than strict JSON for small models, and that tool selection degrades above ~20 entries unless you pre-filter via RAG or a shortlist. Demonstrates safety failures: some models refuse direct harmful asks but comply when roleplayed, so external guards remain necessary.
AI-written summary · about
chapters (12)
- 0:00MCP and tool calling intro
- 1:07MCP tool communication standard
- 2:15Prompting small models for tools
- 3:46Meta’s JSON tool instructions
- 4:28Llama/Qwen tool formats
- 5:35Universal tool calling templates
- 6:38Concise MCP server descriptions
- 7:41Tool selection vs context length
- 8:45Security risks in small models
- 9:27Prompt injection in MCP servers
- 10:30Testing ethical boundaries
- 11:12Conclusion and resources
related talks
34mLLMs, Meet Your Tools: The Magic of Model Context Protocol
43mLLMs, Meet Your Tools: The Magic of Model Context Protocol
29mMCPs are Boring (or: Why we are losing the Sparkle of LLMs)
32mMCP for dummies: Feed more context to LLM
41mBending a Public MCP Server Without Breaking It
50mBeyond local tools: Dive into the Model Context Protocol (MCP)
25mTools in your AI's Toolbox : An introduction to MCP Servers
31mWriting Model Content Protocol (MCP) Servers in Clojure