codexproof

glossary / MCP (Model Context Protocol)

Anthropic's open protocol (late 2024) for connecting AI agents to tools and data - JSON-RPC over defined transports, with servers exposing tools, resources, and prompts that any MCP-speaking client can discover and call.

thin waistgatewaya2aaip
MCP (Model Context Protocol) - One of four doors above the neck - two tools, and deliberately not the building

One of four doors above the neck - two tools, and deliberately not the building.

Hand-authored SVG for this glossary, 2026. Surfaces per main.rs routes; the calibration word "candidate" per the paper's own claim discipline.

Etymology and backstory

born from the N-times-M integration problem - every assistant wiring bespoke connectors to every tool. MCP standardized the socket; the ecosystem effect was immediate and industry-wide, with the spec evolving fast (the paper cites the 2026-07-28 revision; session semantics were retired along the way - one of the terminology updates on the Aug-strata list).

Ecosystem

thousands of MCP servers, client support across major assistants and IDEs, and Eric's own consulting lane - hub-level MCP architecture is precisely the expertise the interview ladder is testing.

In codexproof

one route on the HTTP router (POST /mcp, main.rs line 84) - a hand-rolled JSON-RPC binding, no SDK, exposing exactly two tools: verify and capture. The vector-query fan-out is deliberately NOT an MCP tool (interrogation is HTTP-only) - the Q17 inversion: MCP is the courtesy door for check-and-store so agent stacks integrate with zero new client code, and the waist stays protocol-neutral.

"We treat MCP as one binding, not the architecture - two tools, hand-rolled, at one route - because a thin waist that only works through one protocol above it would be a socket, not a waist."
References: Anthropic, Model Context Protocol specification (the paper's mcp citation, 2026-07-28 revision).

All terms · Questions & answers · The verifier