Anthropic Just Opened the AI App Store: What the Claude Plugin Portal Means for Designers

Anthropic quietly opened a developer portal yesterday. Anyone on a paid Claude plan can now submit MCP connectors and plugin bundles directly to the Claude directory at claude.ai/directory/manage. No partner program required. No Anthropic approval needed to start.

The reaction from the developer community has been mostly about the mechanics: how submissions work, what the review process looks like, what the rate limits are. All useful stuff. But the bigger story for designers and builders is what this portal actually represents.

MCP, the Model Context Protocol, is becoming the connective tissue between AI tools and everything else. Zoom, Templafy, PropellerAds, and dozens of other companies already have connectors in Claude's directory. The AlphaSignal newsletter claims MCP usage is up 110x year over year. That number doesn't appear in Anthropic's official documentation, but even at half that growth, the direction is obvious.

This isn't a developer tools announcement. This is the opening of an app store, and the apps are connectors, plugins, and skills that run inside a conversation.

Image: Claude Marketplace, official Anthropic page, used for editorial commentary

Why most coverage misses the design angle

Most of the coverage I've seen treats this as infrastructure news. Developers get a new submission flow. Security teams get automated scanning. Companies get a distribution channel. All true, and all missing the point for anyone who designs products.

The design question nobody is asking: what does your tool's interface look like when it shows up inside a Claude conversation instead of as a standalone app?

When someone installs your MCP connector or plugin, they're not opening your app. They're talking to Claude and Claude is calling your tool in the background. The interface is the conversation itself. Anthropic is building toward MCP Apps, which are interfaces that render inside the chat. That means every connector and plugin that wants to feel good inside Claude needs design thinking, not just engineering.

This is a new surface area. It's early, it's raw, and it's wide open.

The open question: trust without curation

Here's where I have a real opinion.

The portal has two submission paths. Plugin bundles go through automated validation, a security scan, and then a human reviewer before they go live. MCP connectors go through the same automated scan, but most of them auto-publish as "Community connectors" without a human ever looking at them.

Community connectors can go live after passing an automated scan. The "Verified" label exists, but it follows a separate path. So a user browsing the directory sees a mix of reviewed and auto-published connectors, and the difference isn't always obvious.

This matters more because of what happened two weeks ago. Plugin4Shell, a zero-click remote code execution vulnerability, hit Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI by bypassing SHA pinning on plugins. AIR Security published the technical breakdown on September 17. Anthropic patched Claude Code in version 2.1.179, but the vulnerability demonstrated a real problem: people install extensions without reading their code, and the pins meant to protect them can be bypassed.

The portal's automated scan is the response. It checks for undeclared data sending, validates tool annotations, and scans every commit on the tracked branch. That's better than nothing. But a Community connector that passes an automated scan is not the same as one a human reviewed, and the directory currently makes the distinction less visible than it should.

My take: the open submission model is good. It's how you get a real ecosystem instead of a curated list of partner companies. But trust labels need to be clear enough that a non-technical user can tell the difference between "scanned by automation" and "reviewed by a human." Right now that line is blurrier than it should be.

What designers and builders should actually do

If you build tools, integrations, or workflows, this is a distribution opportunity that's still early.

1. Submit your MCP connector now. The directory is growing fast. Getting in early means visibility before the catalog fills up. The requirements are specific but manageable: remote HTTPS endpoint, OAuth if you act on user accounts, tool annotations for every tool, documentation URL, privacy policy, and a test account for reviewers.

2. Design the in-conversation experience. Don't just wire up the API. Think about how your tool feels when Claude invokes it. What does the user see? What shows up in the chat? If you're building toward MCP Apps, the interface inside the conversation is where your design skills matter most.

3. Audit your installed plugins. If you use Claude Code or any AI coding agent, Plugin4Shell showed that the extensions you trust can be a vector. Check what you have installed, update to patched versions, and remove anything you don't actively use.

4. Build with distribution in mind. If you're building AI tools or agents, the Claude directory is now a channel. A tool that connects to Claude through MCP can go from zero to installed by paying users in days, not months. If you're building on Base44, the backend functions and agent capabilities you already have can be exposed as MCP tools with the right server setup.

The bigger picture

What Anthropic opened yesterday is not just a portal. It's the infrastructure for a marketplace that runs inside a conversation. Every connector added makes Claude more useful. Every plugin extends what it can do. The directory, the review process, and the usage stats are the scaffolding of an ecosystem.

For designers, the meaningful part is this: the next wave of product interfaces won't be standalone apps. They'll be tools that get called inside an AI conversation, and those tools need to feel like they belong there. That's a design problem, and it's one worth solving early.

The portals that look boring on the surface are usually the ones that matter most.