MCP becomes AI agent standard as industry rallies behind it
The gist
MCP has skyrocketed from Anthropic’s internal project to the AI agent integration standard, now backed by OpenAI, Google, and the Linux Foundation.
What to know
- By May–September 2026, MCP evolved from an Anthropic side project to an industry standard, with OpenAI and Google joining the charge.
- The protocol ditched custom connectors for a shared, OAuth 2.1-secured JSON-RPC model, fueling over 10,000 published servers and 97 million monthly SDK downloads.
- MCP’s donation to the Linux Foundation’s Agentic AI Foundation in December 2025 sealed its status as a vendor-neutral backbone for enterprise-scale AI agents.
From Side Project to Standard
MCP's rapid transformation from Anthropic’s internal tool to a vendor-neutral industry backbone was marked by explosive adoption and high-profile institutional support, turning competitors into collaborators within a single year.
The inflection was visible by spring 2026 because MCP no longer looked like an Anthropic curiosity; it looked like an industry standard in formation. Observers who had tracked it since late 2024 described a startling arc: what began as Anthropic engineers’ internal tooling and, later, “just a little spec document, a few SDKs,” mostly Claude-written, local-only, and little more than tools, had, within about a year, become something competitors were embracing, with OpenAI and Google publicly supporting it and companies “releas[ing] MCP servers by the week.”
What turned that momentum into a governance inflection was institutionalization: by December 2025, MCP “was donated to the Linux Foundation’s new Agentic AI Foundation,” with backing from Amazon, Microsoft, Google, OpenAI, and Anthropic, removing the sense that one vendor controlled the protocol. By mid-to-late 2026, that open-standard posture was matched by scale, as analysts called MCP “the default integration standard for AI agents,” citing “over 10,000 published servers and 97 million monthly SDK downloads” as proof it had crossed from promising protocol to broadly adopted industry infrastructure—even as speakers describing the same trajectory framed the later state as “we’re now like at 110 million” downloads, underscoring how far it had moved from its nascent beginnings.
Interoperability—and Its Security Gap
While MCP unified agent integration with a standardized protocol and slashed integration overhead, a persistent security lag—only 8.5% of servers using OAuth—left most deployments vulnerable to tool poisoning and static secret leaks.
MCP replaced one-off connectors with a common client-server protocol built on standardized JSON-RPC messages for tools, resources, and prompts. For most of 2024, every agentic harness invented its own way to attach the same set of capabilities, but MCP let developers expose a capability once and reuse it across environments, turning model-to-tool wiring into a shared interoperability layer.
That architecture scaled because it reduced combinatorial integration work and improved operational tractability: Anthropic shipped the spec in late 2024, and by the end of 2025 every serious agent harness had adopted some version of it, helping drive over 97 million monthly SDK downloads, over 8 million server downloads, and over 10,000 published servers. The appeal was not just convenience but reliability, since MCP was more unit-testable and gateway-friendly than bespoke agent-to-agent wiring.
Security was the other half of the replacement story: MCP externalized credentials and authorization into servers, gateways, and standardized OAuth-based flows instead of leaving teams to roll their own secrets handling inside prompts or app code. But the need for standardization was visible in the gap between design and practice—Only 8.5% of MCP servers use OAuth, 53% of publicly available MCP servers rely on insecure static secrets, and the MCPTox benchmark demonstrated a 72.8% success rate for tool poisoning attacks—which is precisely why a unified, governable protocol mattered for safe orchestration.
