Foundations

Background reading for the local-vs-remote MCP experiment. The writeups (docs/writeup/) report what we measured; these documents explain the concepts the measurements depend on, so a reader can follow the results without prior MCP internals knowledge.

Read in this order:

1. MCP transports — what stdio and streamable HTTP are, how they are deployed, and what is identical across them.
2. Tool discovery and deferral — how an agent learns which tools exist, the ToolSearch deferred-tool mechanism, and why it is a load-bearing experimental variable (it produced a false security signal in one of our runs).
3. Experiment design — arms, tiers, paired seeds, metrics, the two hypotheses, and why there are two experiments.
4. Threat models — the security framing: what each transport exposes and what is protocol-level and therefore transport-agnostic.

The findings these support:

• Visual brief: ../writeup/visual-brief.md
• Long-form writeup: ../writeup/long-form.md
• Slide deck: ../presentation.html

Key external references

• Claude Code: Scale with MCP tool search — the ENABLE_TOOL_SEARCH / deferred-tool mechanism.
• MCP security best practices — confused deputy, scope minimization, and more, with mitigations.
• Model Context Protocol — the protocol itself.