Clerk MCP Server

The Clerk MCP Server is an official, Clerk-hosted Model Context Protocol server that, per the provider, supplies AI agents such as Claude, Cursor, and GitHub Copilot with up-to-date Clerk SDK snippets and implementation patterns. It does not offer generic user management at runtime; rather, it targets developers who write code with Clerk libraries. The server answers questions like: "Which hook do I use to query a logged-in user in React?" or "How do I protect a route in Next.js with Clerk?"

The server is currently in beta according to Clerk's documentation. Functionality may change before general availability.

Official product, repository, and scope

The primary documentation for the Clerk MCP Server is available at clerk.com/docs/guides/ai/mcp/clerk-mcp-server. The GitHub repository clerk/mcp-tools contains source code for Clerk's own MCP tools and integrations — including library functions for client-side and server-side MCP implementations with Clerk auth. Immediately before this entry was created on September 8, 2026, the GitHub API reported exactly 48 stars for the mcp-tools repository. That figure belongs to the entire repository, which encompasses both the hosted server and the library for building custom MCP integrations. Stars are a point-in-time popularity signal, not evidence of quality or security.

The hosted server endpoint is https://mcp.clerk.com/mcp and exclusively supports Streamable HTTP transport — SSE is not supported per the official FAQ. This means a client must be capable of speaking Streamable HTTP; older SSE-only clients cannot use this server.

Available tools and scope

The Clerk MCP Server exposes two tools according to the documentation:

clerk_sdk_snippet — Returns SDK code snippets and patterns for specific Clerk features, such as useUser, currentUser, middleware, organizations, or custom flows.

list_clerk_sdk_snippets — Lists all available snippets and bundle groups.

These tools are intended for the development phase, not production operation. They help apply integration patterns correctly and reduce outdated or incorrect code suggestions from an AI model that would otherwise fall back on training data.

Available snippet bundles

Per Clerk's documentation, the following bundles are available for comprehensive implementations: b2b-saas (complete B2B SaaS setup with organizations, billing, and role-based access), waitlist (waitlist and early-access flows), auth-basics (core authentication hooks), custom-flows (custom sign-in/sign-up flows), organizations (multi-tenant management), and server-side (server-side auth patterns).

Setup: Clerk CLI or manual configuration

The quickest path is the Clerk CLI command clerk mcp install. It auto-detects supported AI clients and registers the server in their configuration so it is available across every project. For Claude Code specifically, the direct command is: claude mcp add clerk --transport http https://mcp.clerk.com/mcp. Restart the client after installation.

For clients the CLI does not support, manual configuration via an MCP JSON file is possible. The documentation covers Claude Code, Claude Desktop, Cursor, Codex, VSCode, Windsurf, and Zed.

Security, limits, and SDK versions

The Clerk MCP Server sends code snippets to the connected AI client, which may forward this information to a model provider depending on its architecture. Because it is a remote server under Clerk's control, the data flow crosses the network. The following points deserve attention:

SDK versions: The snippets provided reflect the current Clerk SDK. Anyone using an older SDK version should always verify patterns against their installed version. AI agents may apply snippets uncritically even when an API has changed between versions.

Model path: Results from this server are passed to the AI client and may appear in the model's prompt or context. Clerk snippets themselves contain no secrets, but a development session's context can include sensitive information that gets processed alongside tool results. Review the data policies of your client and model provider separately.

Untrusted input: Content from external repositories, issue comments, or websites processed by an agent in parallel can affect how Clerk snippets are interpreted (prompt injection risk). The Clerk MCP Server itself is a controlled source, but it cannot govern which other sources an agent reads concurrently.

Beta status: Because the server is in beta, tool names, bundle identifiers, or the endpoint may change without prior notice.

Frequently asked questions (FAQ)

When should I use the Clerk MCP Server? According to Clerk, the provider recommends it for everyone actively building with Clerk so that AI agents always use current patterns rather than outdated training data.

Why do I see an internal server error? This usually indicates a network or transport issue per the FAQ. Disconnect, reconnect, and ensure the client supports Streamable HTTP transport.

Does the server support SSE? No. Only Streamable HTTP is available at https://mcp.clerk.com/mcp.

An MCP-capable client with Streamable HTTP support (e.g. Claude Code, Cursor, Codex); optionally the Clerk CLI for the quickest setup.

Quickest path: run clerk mcp install (auto-detects supported clients). For Claude Code directly: claude mcp add clerk --transport http https://mcp.clerk.com/mcp. Restart the client after installation. For other clients, add the server manually to the MCP JSON configuration.

clerk mcp install

The endpoint https://mcp.clerk.com/mcp is accessible without an API key via Streamable HTTP. The snippet tools do not require a Clerk account.

The server delivers read-only SDK information only. It has no access to Clerk accounts, user data, or production systems.

Clerk processes tool calls on its servers and returns snippet text. The connected AI client may send this content to its model provider; review that provider's privacy policy separately.

SDK snippets may target a different Clerk version than the one installed. Beta status means potential breaking changes. Results may appear in the agent context alongside untrusted content from other sources (prompt injection risk). Account for the data path over the network and model provider.