Anthropic Opens Claude Plugin Portal With 110x MCP Growth This Year

Anthropic has opened a submission portal that gives third-party Claude extensions a structured release path. Developers can package Model Context Protocol (MCP) connectors and Agent Skills as plugins, run automated checks, submit them for human review, choose when to publish, and track discovery and install metrics after launch. Anthropic outlines the changes in its plugin announcement.

Anthropic says MCP usage across Claude products has grown 110-fold this year. Plugins will serve as the primary distribution format for extensions across claude.ai, the desktop and mobile apps, Cowork, and Claude Code.

One package, several capabilities

A plugin can contain an MCP connector, an Agent Skill, or both. Claude Code plugins can also include development-specific components such as commands, hooks, agents, and language server integrations.

Developers can submit a single remote MCP connector by providing its server details. A bundled plugin requires a GitHub repository containing its manifest, MCP configuration, Skills, and any Claude Code components.

A visible review queue

The submission portal is available to developers on paid Claude plans, with no partner program required. Each submission moves through automated validation, a safety scan, and human review.

• Immediate validation: The portal checks the package structure and reports submission errors.
• Safety scan results: Developers can inspect scan output and recommended changes.
• Status tracking: The dashboard shows where a plugin sits in the review process.
• Controlled publication: Approval does not publish the listing automatically. The developer chooses when it goes live.

Human review means publication is not immediate, and Anthropic does not promise a fixed review time. Teams should include that delay in release plans.

Local checks before submission

Claude Code includes three commands for validating and testing a plugin before it reaches the portal:

• claude plugin validate: Checks the manifest and component files for structural problems.
• claude --plugin-dir ./path: Loads a plugin from a local directory so its Skills and commands can be tested before the repository is published.
• claude plugin eval: Runs developer-authored evaluation cases and compares Claude’s results with and without the plugin.

Plugin evaluations make real model calls and count against the developer’s Claude plan. Evaluation suites should cover expected behavior, failure cases, permissions, and responses when an external service is unavailable.

Metrics expose discovery gaps

After publication, the portal reports install counts by Claude product and plugin version. It also shows listing views and the search queries that led users to the plugin.

Those metrics help identify two parts of the distribution funnel. Low listing traffic points to a discovery problem, while strong traffic paired with few installs points to the listing, positioning, permissions, or setup flow. Install counts do not reveal whether people use the plugin after installation, so developers still need telemetry in their own services for task-level usage and retention.

Version-level data can guide support decisions for bundled plugins by showing how quickly users adopt a release. Remote MCP servers require additional care because server-side changes may occur independently of the installed plugin version.

Richer interfaces and managed auth

Claude’s updated MCP support adds two extensions with direct product implications:

• MCP Apps: Connectors can render interactive interface components inside a conversation, allowing users to inspect and manipulate structured results in chat.
• Enterprise Managed Auth: Organizations can configure OAuth centrally, reducing individual setup for eligible enterprise users.

Stateless MCP operation can also simplify horizontally scaled servers by removing the requirement to retain a long-lived protocol session for every connected user. Developers may still need persistent application state for authentication, jobs, user data, and external service activity.

One directory, uneven runtimes

Existing connector listings will continue to work, so current authors do not need to migrate immediately. Anthropic says developers will later be able to convert those listings into plugins.

Anthropic also plans to unify plugin discovery across Claude and Claude Code over the coming weeks. The products currently use largely separate discovery and installation flows, so a shared directory should reduce duplicate distribution work.

Runtime support still varies by product. Chat, Cowork, and Claude Code can load different parts of the same plugin, and Claude Code alone supports components such as hooks, commands, agents, and language server integrations. Developers should test every target surface and consult the platform-support matrix in the build guide.

The new workflow gives MCP server and Skill authors a single package format, a review process, cross-product distribution, and install analytics. Shipping reliably still requires per-surface testing, time for human review, and service-side instrumentation when post-install usage matters.