The Analytics Power Hour #288: Our LLM Suggested We Chat about MCP. Kinda' Meta, No?
23 snips
Jan 6, 2026 Sam Redfern, a Staff Data Scientist at Canva with experience at Meta, dives into the intricacies of Model Context Protocol (MCP). He explains how MCP works like ‘fingers’ for AI models, aiding their access to tools. The conversation highlights the differences between MCP and APIs, the importance of standardization, and the potential for custom implementations tailored to organizations. Sam discusses practical applications at Canva and emphasizes the need to avoid context pollution while navigating the governance risks associated with powerful AI tools.
AI Snips
Chapters
Transcript
Episode notes
LLMs Need 'Fingers' To Act
- MCP gives LLMs “fingers” to touch external data and tools, bridging non-deterministic models and deterministic systems.
- Sam Redfern argues tool use is the core benefit, not the branding or permanent standardization.
Describe Tools Clearly And Narrow Inputs
- Design MCP tools as simple, well-described primitives (name, description, inputs) so agents know how to use them.
- Constrain tool inputs and outputs to reduce mistakes and improve reliable agent behavior.
MCP Is An Early, Evolving Format
- MCP resembles early document standards like XML: a necessary but imperfect bridging format that will evolve.
- Non-determinism in LLMs reduces pressure for rigid syntax compared with older internet standards.