MCPs are Boring (or: Why we are losing the Sparkle of LLMs)
summary
Argues MCP/tool-calling is stalling LLM potential and shows concrete fixes: pass chat history/memories/files into tool calls, offer editable tool inputs and editable tool outputs UIs, and treat code eval as a first-class tool so models generate and register functions. Demonstrates a prototype JS sandbox that auto-creates REST endpoints and a working CRM, cutting dozens of tool calls and minutes of wait to a few seconds and saving token cost — practical demos and measured gains.
AI-written summary · about
chapters (12)
- 0:00Introduction to AI coding
- 1:01The promise and limits of MCP
- 3:46Constraints of tool usage
- 7:15Extending MCP functionality
- 9:31Human-in-the-loop interaction
- 12:04LLMs as creative code engines
- 14:12The power of eval-based agents
- 16:22Iterative tool development
- 19:01Libraries replace tools
- 20:58Building rich UIs for agents
- 22:45Prototype: The JS sandbox
- 26:46Maintaining the magic
related talks
34mLLMs, Meet Your Tools: The Magic of Model Context Protocol
43mLLMs, Meet Your Tools: The Magic of Model Context Protocol
12mPrompting LLMs to use Tools and Resources with MCP
32mMCP for dummies: Feed more context to LLM
32m30 minutes to understand MCP (Model Context Protocol)
01h 31mBuild AI-Powered Automation Tools: MCP Workshop for Developers & Engineers
55mYour MCP Server is Bad (and you should feel bad)
27mYour MCP Server is Bad (and You Should Feel Bad)