MCP Wins, But the War Just Started: Why Standardization Won't Save Us from Lock-in

Source: Dev.to

The Shift Toward Standardization

The chaotic AI industry—where each company locked users in with proprietary standards—has moved toward a unified “standardization of Agentic AI.” MCP’s momentum is now unstoppable:

• 10,000+ active public MCP servers
• 97 M+ monthly downloads of Python and TypeScript SDKs
• Full adoption by major platforms: ChatGPT, Cursor, Gemini, Microsoft Copilot, VS Code, and virtually all developer tools now support MCP
• Immediate infrastructure support: AWS, Cloudflare, Google Cloud, and Microsoft Azure provide deployment environments for MCP servers

MCP is no longer “Anthropic’s convenient standard”; it has become the de‑facto plumbing for running AI agents.

Vendor Neutrality

The donation’s greatest significance lies in establishing vendor neutrality. Until now, MCP, even when open, was perceived as “Anthropic’s standard.” Betting on a competitor’s technology that could change specs at any time posed a risk for other companies. Anthropic explicitly stated the reason for the donation: “to ensure that MCP remains open source, community‑led, and vendor‑neutral.”

Additional Projects Donated to AAIF

AGENTS.md (donated by OpenAI)

A metadata standard for describing tools and capabilities used by agents. It unifies the previously disparate “instructions for agents” across companies.

goose (donated by Block)

A framework for robust AI‑agent development.

Managing these projects under the same foundation is expected to dramatically improve interoperability. A future where “an agent built with goose reads the specs defined in AGENTS.md and executes tools via MCP” is becoming a reality, even with a mix of vendors.

Recent Specification Updates (Nov 25)

• Asynchronous operations – agents can proceed without waiting for time‑consuming processes.
• Statelessness – reduces server‑side state‑management burden, improving scalability.
• Server identity – enhanced verification of connection legitimacy.
• Official extensions – a safe mechanism to extend standard features.

The Claude directory already lists over 75 connectors, and features like “Tool Search” and “Programmatic Tool Calling” via API are becoming robust, moving us toward a “plug‑in‑and‑it‑works” world.

The Ongoing Challenges

Despite the progress, the broader AI development landscape remains fraught with issues:

• Proliferation of standards that become obsolete within months.
• Vendor lock‑in through proprietary “Skills” and “Ecosystems” that depend on specific execution environments (e.g., Claude’s ecosystem).
• Siloization of troublesome operations—even if any tool can connect via MCP, the “Brain (Skills)” needed to use those tools wisely is described in each company’s proprietary specs.

The “Agent‑first” paradigm shift has filled vendors’ moats, reducing learning costs for users but pushing vendors to tie users down with value‑adds like Skills and Ecosystems. Their desperation stems from the fear of becoming replaceable commodities.

Outlook

As Ilya Sutskever noted in his podcast, “We’re moving from the age of scaling to the age of research.” We remain squarely in the “Age of Research.” The pain points—prompt engineering, model hallucinations, non‑deterministic behavior—will persist.

Conclusion: MCP has won the standardization war, but the underlying challenges of AI development and vendor lock‑in remain. The battle is far from over.