agents-mcp
MCP (Model Context Protocol)
Use this skill to decide whether MCP is the right abstraction, configure existing servers in Claude Code or Codex, or build a narrow custom server when repeated agent workflows justify it.
Protocol baseline: https://modelcontextprotocol.io/specification/2025-11-25 (stable). The 2026-07-28 spec finalizes on 2026-07-28 (blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate) — as of this writing that is under three weeks out; treat any new MCP work started now as needing a post-launch compatibility pass, not as a distant future concern. The RC is a fundamental redesign, not an increment: the protocol becomes stateless (the initialize/notifications/initialized handshake and Mcp-Session-Id header are removed, replaced by per-request _meta fields and a mandatory server/discover RPC), Roots/Sampling/Logging are deprecated outright (not just folded into MRTR — migrate to tool-parameter directories, direct provider API calls, and stderr/OTel logging respectively), and Elicitation moves to the Multi Round-Trip Requests (MRTR) pattern. TypeScript SDK v2 (beta as of 2026-07) targets this spec and ships stable alongside it; v1.x stays the supported production lane for at least 6 months after v2 stabilizes. Build to the stable 2025-11-25 spec today; plan a scoped migration window once 2026-07-28 finalizes rather than pre-adopting RC shapes in production.
Governance (factor into vendor-trust judgment, not just the tech): Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation, effective 2025-12-09 (blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation; AAIF founding members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI). The existing maintainers keep full technical authority over the spec via the SEP process — the foundation explicitly "will not dictate the technical direction of MCP" — so day-to-day spec/SDK guidance in this skill is unaffected. What changes for your judgment: MCP is no longer a single-vendor bet, which lowers long-term protocol-abandonment risk and is a legitimate factor when a customer or security review asks "who owns this protocol" — cite the foundation, not Anthropic, when that question comes up.
Quick Reference
- use MCP when you need reuse, explicit permissions, discovery, or a stable tool contract across sessions
- prefer an existing official or vendor-maintained server before building custom infrastructure
- default to a low-cost health check before trusting a server
- build the narrowest possible tool surface and keep write access gated