Building an MCP server in 2026: what's changed since launch
The Model Context Protocol went from a single vendor's launch to a Linux Foundation standard with 10,000+ live servers. Here's what that means for building one today.
Jordan Lee
AI Correspondent
The Model Context Protocol (MCP) — a standard way for AI assistants to call out to external tools and data sources — has grown fast since its launch. Monthly SDK downloads reached 97 million by March 2026, up from around 100,000 at launch, and the official MCP Registry listed roughly 9,652 latest server records (28,959 including versions) as of late May 2026. By mid-2026, more than 10,000 active public MCP servers were in operation, with first-class client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and VS Code.
From one vendor to a standard
In December 2025, Anthropic — which originally created MCP — donated it to the Agentic AI Foundation under the Linux Foundation, moving it from a vendor-led project to a vendor-neutral, community-governed standard in under two years. The most recent specification, dated 2026-07-28, is a significant architectural shift: it moves MCP to a stateless core that removes transport-level session management entirely, so servers can scale on ordinary HTTP load-balanced infrastructure instead of needing sticky sessions.
What building one looks like now
Practically, an MCP server is still a thin wrapper: you define a set of tools (functions with typed inputs and descriptions) and, optionally, resources (data the model can read), and expose them over the protocol's transport. The 2026-07-28 spec's move to statelessness is the main thing worth designing around today — building a server that doesn't assume a persistent connection to a single client keeps it compatible with load-balanced, serverless-style deployment rather than requiring a dedicated long-lived process per client.
Adoption data suggests coding tools remain the highest-volume early use case — at the observability company Honeycomb, for instance, nearly 20% of monthly interactive queries are now made by agents rather than humans directly. That's a reasonable signal of where MCP servers currently pay off fastest: internal developer tooling and data sources that coding agents need to query.
More from Guides
A practical guide to RAG that actually works in production
Beyond the tutorial: chunking strategy, retrieval evaluation, and the failure modes that only show up at scale.
When fine-tuning beats prompting (and when it doesn't)
A cost-and-quality framework for deciding between prompt engineering, RAG, and fine-tuning for a given use case.