MCP started as an Anthropic project in late 2024. By late 2025 it had become a vendor-neutral open standard under Linux Foundation governance, with adoption by OpenAI, Microsoft, AWS, Google, and a range of enterprise software vendors.
The Linux Foundation Donation
In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. This is the same organizational model that governs Kubernetes, Node.js, and dozens of other foundational technologies — a neutral home where no single company controls the roadmap.
The practical implications:
- No lock-in to Anthropic’s products. The protocol belongs to the foundation, not to Claude. Any AI vendor can implement it without Anthropic’s permission or involvement.
- Stable governance. The Linux Foundation has a track record of maintaining projects through commercial turbulence. Foundation projects don’t get deprecated because a company pivots.
- Broader contribution. Governance under a foundation means other companies can contribute to the core protocol specification, not just implement it.
For long-term infrastructure decisions, the foundation status removes the category of “what if Anthropic changes direction” risk.
Corporate Adoption
Corporate adoption as of early 2026:
| Company | Integration |
|---|---|
| OpenAI | Adopted March 2025; integrated across Agents SDK and ChatGPT desktop |
| Microsoft | Windows 11 MCP integration announced at Build 2025; Azure MCP Server |
| AWS | Official awslabs/mcp server collection |
| GenAI Toolbox for BigQuery, Cloud SQL, Spanner, AlloyDB | |
| Cloudflare | Official server for Workers, KV, R2, and D1 |
| Atlassian | Official Jira/Confluence MCP server |
| GitHub | Official GitHub MCP server; runs the MCP Registry |
| Block, Bloomberg | Production deployments |
OpenAI’s adoption in March 2025 made MCP cross-vendor. When Claude, ChatGPT, and GitHub Copilot all speak MCP, the protocol is effectively an industry standard rather than a single-vendor approach.
Microsoft’s Windows 11 integration announcement at Build 2025 extended this to the OS layer — MCP support built into the operating system, not just AI applications. The Azure MCP Server brings it into enterprise cloud infrastructure.
Vendor-Maintained Servers
A distinct pattern emerged alongside foundation governance: major platforms began maintaining their own MCP servers rather than relying on the Steering Group or community.
GitHub took ownership of the GitHub MCP server. Brave took ownership of their search server. Snowflake Labs built and maintains the official Snowflake MCP server. Google maintains the GenAI Toolbox for their cloud databases.
This is the healthiest possible outcome for ecosystem quality. A server maintained by the team that builds and operates the underlying platform will track API changes faster, support new features earlier, and reflect the product team’s knowledge of how the API is actually meant to be used. Compare this to a community-maintained server for a vendor’s platform, where the maintainer is guessing at the right behavior and can’t get early access to API changes.
When a vendor-maintained server exists, it generally tracks API changes faster, supports new features earlier, and reflects the product team’s knowledge of correct API use. The Snowflake Labs server and the GitHub-maintained GitHub server are the preferred options over community alternatives for those platforms.
What 97 Million Downloads Means
The MCP Protocol Architecture note covers adoption metrics, but it’s worth framing them in context of governance. As of early 2026:
- 97+ million SDK downloads
- 10,000+ production servers
- 5,800+ community servers in the awesome-mcp-servers directory
- Official registry at registry.modelcontextprotocol.io maintained by GitHub
Block (Square’s parent) and Bloomberg running production MCP deployments indicates the protocol has been tested against enterprise security, compliance, and reliability requirements, not only by early adopters.
Protocol Stability
One practical question for infrastructure investment: is the protocol stable enough to build on? The short answer is yes, with a footnote.
The core protocol — JSON-RPC communication, the three primitives (tools, resources, prompts), stdio and HTTP transports — is stable. Code written against MCP in 2024 still works in 2026. The Python SDK is on v1.x with v2 in development, following semantic versioning with migration paths.
The footnote: the protocol is still evolving. New capabilities get added (OAuth 2.1 for HTTP transport, elicitation for server-initiated user queries, dynamic tool definitions). These are additive — they don’t break existing servers. But if you’re implementing against the spec directly rather than using an SDK, you’ll need to track the changelog. Using an official SDK means the protocol evolution is handled for you.
For data engineering infrastructure, the governance picture is: a foundation-owned standard with broad vendor adoption, a stable core, and an ecosystem of both official and community implementations.