Anthropic Verified バッジはなぜ取れないのか」— 審査を通過する8つの設計原則 Anthropic Verified バッジはなぜ取れないのか」— 審査を通過する8つの設計原則

Many developers who publish MCP (Model Context Protocol) plugins to Anthropic's directory find themselves wondering why the Verified badge never appears next to their listing. A detailed Zenn post tackles this question head-on, mapping out the review pipeline and distilling eight design principles that meaningfully improve a plugin's chances of passing.

Anthropic's review process is believed to combine automated scanning with manual evaluation. The automated layer flags common technical risks: vulnerable dependencies, license conflicts, and hardcoded secrets. The manual layer then assesses softer but equally important criteria — transparency about what data the plugin accesses, resistance to prompt injection, and whether users can reasonably predict the tool's behavior before invoking it. Clearing only one layer is not enough; the design must satisfy both simultaneously.

The eight principles outlined in the article cluster around three themes: security, transparency, and user protection. On the security front, least-privilege scope design and rigorous input validation are highlighted as non-negotiable baselines. Transparency demands clear, jargon-free descriptions of what each tool does and explicit disclosure of the data types it collects or transmits. User protection calls for confirmation flows that prevent unintended side effects and graceful fallback handling when errors occur — so a failure mode never leaves the user in an ambiguous or harmful state.

These standards echo Anthropic's broader Constitutional AI philosophy. Just as the company trains its models to be helpful, harmless, and honest, it appears to extend that same expectation to the tooling that connects to those models. The parallel with OpenAI's GPT Store review policies and Google's emerging verification practices for Gemini tools is striking: across the industry, the bar for agentic tool quality is rising in lockstep with the capabilities of the models themselves.

For developers, the practical takeaway is to treat the Verified badge not as a bureaucratic hurdle but as a useful design checklist. Plugins built to satisfy Anthropic's criteria tend to be more robust, more trustworthy, and likely to retain users over time — outcomes that benefit everyone regardless of badge status. As the MCP ecosystem matures and more plugins compete for visibility, Verified status may well become one of the primary signals users rely on when choosing which tools to grant model access. Building with that standard in mind from day one is likely the most efficient path forward.