---
type: source-digest
topic: ue-mcp-strategy
status: active
created: 2026-06-16
updated: 2026-06-16
source_notes:
  - external
derived_from:
  - web-research
confidence: high
source_url: https://www.anthropic.com/news/model-context-protocol
---
# mcp-spec

## Why This Matters
MCP is the "open tool surface" half of Darsh's build-vs-expose question. To reason about whether UE should *open an MCP* vs *build a harness*, you need to know what MCP actually standardizes (tool discovery for agents) versus what it does not (it adds no new capability the host app didn't already have). The non-obvious point: MCP's value is the protocol/standard, not the plumbing — and a standard is exactly what can commoditize a runtime if it's engine-neutral.

## Key Takeaways
- **Launch:** Anthropic open-sourced MCP on **Nov 25, 2024**, described as "a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments."
- **Created by** David Soria Parra and Justin Spahr-Summers at Anthropic; positioned as an open-source standard from day one.
- **Architecture = host / client / server.** The *host* is the AI app (Claude Desktop, Cursor, an IDE) holding the LLM(s); it spawns one *client* per *server*, each client holding a dedicated 1:1 connection. The *server* exposes capabilities. Host enforces user approval between model and any server action (security boundary).
- **Three server primitives:**
  - **Tools** — model-invokable actions (call an API, run a command, mutate state).
  - **Resources** — readable data/context (files, DB rows, repos) pulled into the model's context.
  - **Prompts** — reusable instruction templates / workflows the server ships.
  - (Client-side primitives also exist: **roots, sampling, elicitation**.)
- **Transport / wire format:** JSON-RPC 2.0 data layer; transports are **stdio** (local, same machine) and **Streamable HTTP** (remote, HTTPS + SSE). Data layer covers lifecycle, **capability negotiation**, and the primitive semantics.
- **Dynamic discovery (the core idea):** instead of hardcoding integrations at training time, the model queries the server at runtime for a catalog of what's available, then picks tools per task. Capability discovery, schema negotiation, and access control are *first-class protocol constructs* — that's the differentiator vs ad-hoc APIs/plugins.
- **Analogy:** widely called the "**USB-C port for AI applications**" — one universal extension point.
- **Pre-built servers at launch:** Google Drive, Slack, GitHub, Git, Postgres, Puppeteer.
- **Early adopters named:** Block, Apollo (integrated); Zed, Replit, Codeium, Sourcegraph (dev-tool partners). Block CTO: "Open technologies like the Model Context Protocol are the bridges that connect AI to real-world applications."
- **Adoption (1-yr milestone, ~Nov 2025):** described as one of the fastest-growing open-source AI projects — ~**97M monthly SDK downloads**, ~**10,000 active servers**, first-class client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, VS Code. **OpenAI and Google DeepMind adopted it** — i.e., it became a cross-vendor standard, not an Anthropic-only thing. [NON-OBVIOUS: a rival-adopted standard is what makes the "commoditization" risk real.]
- **Governance:** Anthropic donated MCP to the **Agentic AI Foundation** (a Linux Foundation directed fund) — neutral stewardship, reinforcing standard-not-product status.

## Entities
- **Anthropic** — creator; **David Soria Parra**, **Justin Spahr-Summers** — authors.
- **OpenAI, Google DeepMind** — adopters (cross-vendor).
- **Agentic AI Foundation / Linux Foundation** — current steward.
- **Hosts/clients:** Claude Desktop, Cursor, Windsurf, VS Code, ChatGPT, Gemini, Copilot.

## Sources (all URLs)
- https://www.anthropic.com/news/model-context-protocol (official announcement, launch date, primitives roles, pre-built servers, adopters)
- https://en.wikipedia.org/wiki/Model_Context_Protocol (creators, definition, Nov 2024)
- https://www.analytical-software.de/en/the-model-context-protocol-mcp-deep-dive-into-structure-and-concepts/ (host/client/server, JSON-RPC, layers)
- https://workos.com/blog/mcp-features-guide (tools/resources/prompts + roots/sampling/elicitation)
- https://blog.modelcontextprotocol.io/tags/announcement/ and https://zuplo.com/blog/one-year-of-mcp (1-yr adoption stats, foundation donation)
- https://newsletter.pragmaticengineer.com/p/mcp ("USB-C port" framing, building-block context)

## Open Questions
- Exact spec revision current as of 2026-06-16 (spec is versioned/dated; "Streamable HTTP" superseded the original SSE transport in 2025) — confirm against modelcontextprotocol.io before quoting wire details authoritatively.
- The ~97M downloads / 10k servers figures come from a project blog milestone, not an audited source — treat as directional.