Model Context Protocol (MCP)

MCP is a wire protocol that lets a model or agent talk to tools and context sources in a standardized way. An MCP server exposes capabilities (read files, query a database, call an API); an MCP client (the model or agent) discovers and invokes them. The spec, reference SDKs in TypeScript / Python / Go / Rust, and a large community-maintained ecosystem of servers all live in the open under MIT. Originally seeded by Anthropic; donated to the Linux Foundation's Agentic AI Foundation (AAIF) in December 2025. MCP matters because protocols at the integration layer end up owning the ecosystem. HTTP, SMTP, BGP, and TCP are the analogous historical examples; their openness allowed the internet to be a platform anyone could build on. The sibling protocols are A2A (Google-seeded, agent-to-agent rather than agent-to-tool; complementary not competing) and ACP (less widely adopted). MCP has become the de facto standard for tool access by agents in 2025-2026: supported by Claude Code natively, by Goose, by OpenHands, by Cursor, and increasingly by hosted services that want to expose themselves as MCP servers. Production-ready and broadly deployed. The interesting open question is whether MCP's evolution stays community-driven now that it sits under AAIF, or whether Anthropic's seeded status continues to influence the spec's direction. So far it has been treated as a published standard rather than a controlled product.