World's Fair 2025
MCPs are Boring (or: Why we are losing the Sparkle of LLMs) - Manuel Odendahl
Overview
This talk argues that the current focus on Multi-Call Protocol (MCP) and tool calling for LLMs is limiting their potential. The speaker contends that LLMs are powerful code generators and language producers, capable of much more than simply calling predefined functions with fixed schemas. The core thesis is that by treating LLMs as sophisticated code-generation engines and embracing recursive creation, developers can unlock greater flexibility and power, moving beyond the "boring" limitations of current tool-calling paradigms.
Who should watch
- AI Engineers
- Product Managers
- Developers building LLM applications
- Those interested in agentic workflows and tool integration
- Anyone feeling constrained by current LLM tool-calling limitations
Key takeaways
- LLMs excel at generating code, which can be used to create dynamic tools and functionalities on the fly, rather than relying on static, predefined schemas.
- The current MCP and tool-calling approach can be inefficient, leading to large token usage, slow responses, and difficulty in handling complex queries or numerous tools.
- Leveraging LLMs' code generation capabilities allows for the creation of richer, more interactive tools, including UIs for editing inputs and outputs before they are processed.
- The concept of `eval` can be powerful, enabling LLMs to interpret and execute code (like SQL queries) based on provided schemas, leading to more deterministic and efficient results.
- Developers should think recursively: ask the LLM to write code to do something, then ask it to write code to write that code, creating layers of abstraction and capability.
- Instead of rigid tool calls, consider exposing libraries or APIs that LLMs can interact with by generating the necessary code, offering more flexibility.
- The "sparkle" of LLMs is being lost by focusing on restrictive tool-calling mechanisms; embracing their generative power can restore this magic.
Notable quotes
*MCPs are boring, and we are losing the sparkle of LLMs.*
*LLMs are language producers. They can create every word under the sun.*
*Why are we stuck with like tools that only that don't even work that well?*
Unofficial community note. Prefer the recording for nuance.