Register Local Custom MCP Servers

A Custom Local MCP server is a privately developed package (Python or npm) stored securely in a local JFrog Artifactory repository. It provides specific tools and capabilities to AI agents — for example, connectors to internal databases, internal APIs, or private platforms. Once registered in the JFrog MCP Registry, the server becomes discoverable in the AI Catalog, allowing developers to securely connect their AI agents to it via the governed JFrog Agent Guard.

Registration acts as the approval mechanism for these servers within your organization and links the physical Artifactory binary to the MCP Registry. You can register local servers through the JFrog Platform UI or the AI Catalog REST API.

Prerequisites Before Registering a Local Custom MCP server

Ensure the following requirements are met:

• Upload the binary: Publish your MCP package to a local npm or local PyPI repository using the standard package managers. All uploaded artifacts must be fully self-contained and include all runtime dependencies.
• Check project scoping: Ensure the local repository is assigned to the JFrog Project where you intend to register the server.
• ALLOW permission for the target project within the AI Catalog.
• READ access to the repository containing the custom MCP server package.
• Custom MCP registration currently supports npm and PyPI; Docker-based MCPs are available through Discovery (public catalog) and can be allowed for your project from there.

Register a Local Custom MCP Server

📘

Prerequisites

• Upload the binary: Publish your MCP package to a local npm or local PyPI repository using the standard package managers. All uploaded artifacts must be fully self-contained and include all runtime dependencies.
• Check project scoping: Ensure the local repository is assigned to the JFrog Project where you intend to register the server.
• ALLOW permission for the target project within the AI Catalog.
• READ access to the repository containing the custom MCP server package.
• Custom MCP registration currently supports npm and PyPI.
• Public MCPs are available through Discovery (public catalog) and can be allowed for your project from there.

Register a Custom MCP Server

Registering an MCP server links the physical Artifactory binary to the MCP Registry, approving it for organizational use. This can be done via the UI or REST API.

Register through the UI

1. In the Platform module, navigate to AI/ML > Registry.

2. Select the project for which you want to register the MCP server.

3. Click (+), and select Upload Custom MCP to open the registration pane.

4. Select the local repository that contains the server binary, and enter the path to the artifact.

5. Define the server Name: enter a unique identifier for the server.

   • The registry automatically formats the display name as <type>:<repository>:<name>.

6. (Optional) Configure runtime settings:

   • Runtime Arguments: Enter the required strings for execution and provide a description for each argument. These are optional command-line execution flags.
   • Environment Variables: Define required keys with descriptions and toggles for isSecret (to mask values in the UI) and isRequired.

7. Enter the path to the README file in Markdown format.

8. Define the Tool Policy: choose how to govern the tools within this server:

   • Allow all tools: Automatically approves all current tools and any tools added to this server in future updates.
   • Select tools manually: Opens the Allow List and Deny List configuration. This enables you to define granular policies using static text and Regex patterns. See Configure Tool Policies.

9. Click Complete Registration. The MCP server details are displayed in the Registry page.

Register through the REST API

To automate custom MCP registration in your CI/CD pipeline, send a POST request to the registration endpoint:

https://<JFROG_DOMAIN>/ml/core/api/v1/mcp-registry/custom-server.

Example 1: With runtime arguments and environment variables

curl --location 'https://{JFROG_DOMAIN}/ml/core/api/v1/mcp-registry/custom-server' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer {TOKEN}' \
--data '{
    "projectKey": "ai-research",
    "repoArtifactInfo": {
      "repoKey": "ai-pypi-local",
      "repoPath": "org/project/mcp-server/1.0.0/mcp_server-1.0.0.tar.gz",
      "readmePath": "org/project/mcp-server/1.0.0/README.md"
    },
    "mcpServerInfo": {
      "description": "Internal Database Connector for AI Agents",
      "version": "1.0.0",
      "environmentVariables": [
        {
          "name": "DB_CONNECTION_STRING",
          "description": "Connection string for the production DB",
          "format": "string",
          "isRequired": true,
          "isSecret": true
        }
      ],
      "runtimeArguments": [
        {
          "name": "--port",
          "description": "Runtime port for the MCP server",
          "format": "integer",
          "type": "int",
          "defaultValue": "8080"
        }
      ]
    }
}'

Example 2: Without runtime arguments and environment variables

curl --location 'https://{JFROG_DOMAIN}/ml/core/api/v1/mcp-registry/custom-server' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer {TOKEN}' \
--data '{
    "projectKey": "ai-research",
    "repoArtifactInfo": {
      "repoKey": "ai-npm-local",
      "repoPath": "org/project/simple-mcp/1.0.0/simple-mcp-1.0.0.tgz",
      "readmePath": "org/project/simple-mcp/1.0.0/README.md"
    },
    "mcpServerInfo": {
      "description": "A basic utility MCP server with default settings",
      "version": "1.0.0",
      "environmentVariables": [],
      "runtimeArguments": []
    }
}'

JSON Payload Parameters

Managing Registrations

Discovery

Registered MCPs are listed in the AI Catalog. You can filter tools by Name or Type and sort them by Name or Last Update.

Unregister Custom MCP Servers

Unregistering an MCP server immediately removes its metadata and governance policies from the AI Catalog, disconnecting it from developer IDEs, but does not delete the physical binary from Artifactory.

Unregister Through the UI

1. In the Platform module, navigate to AI/ML > Registry.

2. Select the project to which the MCP server you want to remove is registered.

3. Locate the MCP server and click Unregister.

4. Confirm the action in the confirmation message.

The MCP server is immediately removed from the MCP Registry and developers can no longer connect to it.

Unregister Through the REST API

Send a DELETE request to the unregister endpoint, passing the server name and project key:

curl -X DELETE "https://<JFROG_URL>/api/v1/ml/registry/mcp/unregister?name=<MCP_NAME>" \
     -H "Authorization: Bearer <TOKEN>" \
     -H "JFrog-Project-Key: <PROJECT_KEY>"

Where:

• <JFROG_URL> – Your JFrog Platform URL.
• <MCP_NAME> – The name of the MCP server to unregister.
• <TOKEN> – Your authentication token.
• <PROJECT_KEY> – The project key for the MCP server.