Google OKF + MCP — the AI context stack as meaning plus action

The video starts with a useful distinction. A modern model can write a query, but it cannot answer a company-specific metrics question unless it has access to the definitions, data, and relationships behind that metric. The missing part is context.

That shifts the design problem from model selection to context architecture.

OKF is presented as a meaning layer: a graph of metrics, customers, docs, and definitions. MCP is the live, governed tool-access layer. One tells the model what things mean; the other lets it act.

The video connects this to Karpathy’s LLM wiki idea, where knowledge compounds because the agent can maintain its own working context.

The stack is useful because either layer alone is incomplete. Knowledge without action produces reports. Tool access without meaning produces risky automation.

For Claude and Codex deployments, the practical sequence is: organize meaning first, then add least-privilege MCP actions.