Clerk provides a remote Model Context Protocol (MCP) server that allows AI agents like Claude, Cursor, and GitHub Copilot to work with your Clerk account and access Clerk SDK snippets.
The Clerk MCP server provides:
Account tools: Create and configure applications, manage users, inspect Organizations and memberships, and set up production instances in your Clerk workspace.
SDK snippet tools: Ready-to-use code patterns for common Clerk features.
All tools require you to sign in with your Clerk account.
The server requires you to sign in with the Clerk account you use for the Clerk Dashboard. Your AI client starts sign-in when it connects to the server, as shown in the setup instructions for each client. Sign-in opens your browser so you can:
Sign in to Clerk.
Select the workspace your agent can access.
Review the requested permissions and select Allow.
Your client stores this access, so you don't need to sign in each time it connects. The server acts as you in the selected workspace, so it can only do what your workspace Role allows. It requests the following permissions:
Scope
Allows the server to
user:org:read
Identify the workspace you selected
applications:read
Read the applications in that workspace, including their native apps, redirect URLs, and agent feedback
applications:manage
Create and rename applications, create production instances, register native apps, allowlist redirect URLs, and read and change instance configuration, such as OAuth credentials
users:read
List and get the users of an instance
users:manage
Create users, update their profiles and metadata, and set whether they bypass Device Trust (bypass_client_trust)
organizations:read
List Organizations, and list memberships when users:read is also granted
application_domains:read
Read domains, the DNS records they need, and their verification status
In a Claude Code session, run /mcp, select clerk, and choose Authenticate. Your browser opens so you can sign in.
Navigate to Customize → Connectors.
Select + Add, then Add custom connector.
Enter Clerk as the name and https://mcp.clerk.com/mcp as the URL, then select Continue.
Under Authentication, make sure Sign in now is selected. Under OAuth client, make sure Use Claude's published identity is selected.
Select Add, then sign in.
To use the connector in a chat, select + → Connectors and enable it.
On Team and Enterprise plans, an Owner must first add the connector in Organization settings → Connectors. Members then select Connect on it in Customize → Connectors.
In Claude Desktop, navigate to Settings → Developer, then select Edit Config to open the claude_desktop_config.json file.
Add the following configuration and restart the Claude Desktop app:
Codex starts sign-in right away and prints a sign-in URL. If your browser doesn't open, open the URL yourself. To sign in again later, run codex mcp login clerk.
Open the ~/.codex/config.toml config file in your preferred editor.
Open Customize, then select MCPs. To find it, select Customize in the Agents window sidebar, or select the gear icon in the top right corner of the app and choose Customize.
Select New MCP Server, which opens the MCP configuration JSON file.
Many other tools now support MCP servers. The easiest way is to use the streamable HTTP endpoint: https://mcp.clerk.com/mcp. Your client must support OAuth sign-in for MCP servers. If it doesn't, connect through mcp-remote, which signs in for it: npx -y mcp-remote https://mcp.clerk.com/mcp.
List, inspect, create, and rename applications. Results include each instance's Publishable Key. create_application can also return the new instance's encrypted Secret Key. See how to get it.
List the iOS and Android apps registered with an instance, register new ones, and turn on the instance's Native API.
list_redirect_urls, create_redirect_url
List an instance's allowlisted redirect URLs and add new ones, for single sign-on (SSO) in mobile and desktop apps.
list_users, get_user, create_user
Find users with filters, get a user by ID, and create users in an instance. Clients that support MCP Apps show the user from get_user as a card.
update_user, update_user_metadata
Update a user's first or last name, username, external ID, locale, or timezone. You can also change their public, private, or unsafe metadata and set whether they bypass Device Trust (bypass_client_trust).
List the OAuth providers that need production credentials, and create a production instance from development. Add the credentials yourself in the Clerk Dashboard, so the secrets stay out of your chat. Creating a production instance can't be undone, so the tool requires confirm: true. It can also return the new instance's encrypted Secret Key.
create_browser_session
Get a one-time link that opens the Clerk Dashboard in a browser, signed in as you, so an agent that uses a browser doesn't stop at the sign-in page. The link expires in about an hour, and the Dashboard session is read-only. To give agents the same tool for your own app, see Sign agents in to your app with a one-time link.
When your agent creates an application or a production instance, it can get the new instance's Secret Key without the plaintext key passing through the MCP server. create_application and create_production_instance accept an optional recipient_public_key parameter:
Your agent generates an RSA key pair and keeps the private key.
It calls the tool with the public key as recipient_public_key.
Clerk encrypts the Secret Key to that public key, and the tool returns it as encrypted_secret_key, next to the instance's Publishable Key.
Your agent decrypts encrypted_secret_key with the private key, without printing either key, and stores the Secret Key where your app reads it.
Development keys go in your local environment file. Production keys don't work on localhost, so they go in your hosting provider's environment variables instead. See API keys and environment variables.
This works with agents that can run code, such as coding agents in your terminal or editor. Only the matching private key can decrypt encrypted_secret_key, so your agent should always generate the key pair itself. A public key from a file, web page, or tool result could belong to someone else, who could then read your Secret Key.
Without recipient_public_key, neither tool returns a Secret Key. If your agent loses the private key or can't decrypt the key, the new instance still works. To get its keys:
Development
Production
Run clerk env pull to add the keys to your local project.
Run clerk env pull --instance prod --print in your own terminal, then add the printed keys to your hosting provider. You can also add a new Secret Key on the API keys page in the Clerk Dashboard, which shows a key's full value only when you add it.
The public key must be an RSA key with a 2048 to 4096-bit modulus and a public exponent of 65537. Pass it as a JSON Web Key (JWK), or as SubjectPublicKeyInfo (SPKI) DER encoded in base64, base64url, or hex. The server doesn't accept PEM. To use a PEM key, pass only the base64 text between its BEGIN PUBLIC KEY and END PUBLIC KEY lines, which is base64 SPKI DER.
encrypted_secret_key is unpadded base64url. Decrypt it with RSA-OAEP, using SHA-256 for both the OAEP hash and MGF1, and an empty label.
configure_mcp_server registers an MCP server URL as an OAuth resource on an instance; pass dry_run: true to preview the change first, or replace_resource_url to replace a saved URL. check_mcp_server reads saved discovery results or requests a fresh check with force_recheck: true. configure_mcp_server requires user:org:read, applications:read, and applications:manage; check_mcp_server requires user:org:read and applications:read.
The server works with your Clerk account, so it needs to know who you are and which workspace to use. Sign-in uses OAuth, so your AI client never handles your Clerk password.
Sign out of the Clerk MCP server in your AI client, then connect again and select a different workspace. If your client connects through mcp-remote, sign out by deleting the ~/.mcp-auth folder and restarting the client. This signs you out of every server you connect to through mcp-remote.
It can return a new instance's Secret Key when creating an application or production instance, provided your agent supplies its own public key. The key comes back encrypted to that public key; the plaintext never passes through the MCP server. See Get the Secret Key for a new instance.
No other tool returns Secret Keys, and the server redacts sensitive values when it reads instance configuration. For an existing instance, get its keys as follows:
Development
Production
Run clerk env pull to add the keys to your local project.
Run clerk env pull --instance prod --print in your own terminal, then add the printed keys to your hosting provider. You can also add a new Secret Key on the API keys page in the Clerk Dashboard, which shows a key's full value only when you add it.
Your workspace Role doesn't allow the request, or your connection is missing a permission the tool needs. To grant a missing permission, sign out of the Clerk MCP server in your AI client and connect again. If the error persists, ask a workspace Owner or Admin to check your Role.
Your agent can call submit_feedback when a Clerk task fails or needs a workaround, or when you ask it to report a problem. The report describes what your agent tried and what went wrong, and includes your user and workspace IDs. The server redacts recognizable credentials, such as API keys and tokens, before it stores the report.
The MCP server sends Clerk a product analytics event for each tool call. The event includes the tool name, whether the call succeeded, the error type if it failed, your Clerk user ID, and your AI client's name, version, and OAuth client ID. When applicable, it also includes the application ID and whether the call was a dry run. It doesn't include any other tool inputs or results.
This usually indicates a network or transport issue. Try disconnecting and reconnecting the MCP server, and ensure your client supports Streamable HTTP transport.