Overview

FusionAuth has a model context protocol (MCP) server which allows you to manipulate a FusionAuth instance using any MCP compatible client.

It does so by exposing each FusionAuth API operation as an MCP tool.

This MCP server is a preview release and may break, change or be discontinued in the future.

You can use this MCP server to interact with FusionAuth using natural language to:

  • perform quick setup of FusionAuth features: “set up a custom registration form for the Pied Piper application. require email address and password, but allow first name and favorite color to be optional. put them all on one page”
  • to explore an existing instance’s configuration: “is PKCE enforced for all my applications”
  • to export configuration as a kickstart or Terraform file: “please export the configuration for the Pied Piper application to a kickstart file”

The MCP server is installed using the stdio transport and so pairs well with FusionAuth’s local instance when run locally.

Using this MCP server provides the MCP client with a FusionAuth API key. Use this MCP server for development and test instances only.

FusionAuth is not responsible for potentially leaking sensitive information.

Prerequisites

  • A running FusionAuth instance
  • Node.js

Configuration

First, set up a limited API key in the FusionAuth instance. Here's documentation on creating an API key and configuring the correct permissions for an API key.

Next, configure your MCP client to use the FusionAuth API MCP server.

For example, to add to Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json to include the fusionauth-mcp-api key below.

If you don't have any previous MCP servers installed, it would look like this:

{
  "mcpServers": {
      "fusionauth-mcp-api": {
      "command": "npx",
      "args": [
        "@fusionauth/mcp-api"
      ],
      "env": {
        "API_KEY_APIKEYAUTH": "<your fusionauth api key>",
        "API_BASE_URL": "http://localhost:9011",
        "USE_TOOLS": "retrieve,search"
      }
    }
  }
}

The USE_TOOLS env variable above essentially restricts the available tools to read-only operations. You can omit this variable to allow full access, but you will need a model that can handle about 200k tokens.

Consult your MCP client's documentation to determine exactly how to add an MCP server to your client.

Example Prompts

  • "which tools do you have access to?" should show you all the FusionAuth API tools
  • "how many fusionauth applications do I have?"
  • "how many users do I have in my fusionauth instance?"
  • "add a user with an email address of test@example.com and a password of 'password'" (requires less restricted tools)

Restricting Tools

The default MCP Server has a tool for every API endpoint of FusionAuth. Over 300 of them! However, the tools, descriptions, requests, and responses combine to nearly 200k tokens, which can exceed the context window of many MCP clients.

You can restrict which tools this MCP server makes available by setting the USE_TOOLS env variable as shown above.

Each tool is defined by its prefix. The default prefixes are:

  • create
  • delete
  • patch
  • update
  • retrieve
  • search

There is also an other tool bucket that contains every tool with another prefix. The all tool bucket includes all tools and is the default value.

For example, if you don't need to use any delete or patch methods, the following setting reduces the tool list by 20%.

USE_TOOLS="create,update,retrieve,search,other"

If you only want to allow read-only operations, use the following configuration:

USE_TOOLS="retrieve,search"

This also reduces the tool list size by 66%.

Securing Your MCP Server

There are three different approaches to secure your MCP server and you should combine them to enable secure access to your FusionAuth instance while still meeting your functionality needs. In order of granularity (from coarse to fine-grained):

  • Point to the correct instance. This tool is in preview and we suggest you only point it at a local, dev instance.
  • Enable the correct type of tools. In addition to reducing context window usage, limiting the type of requests your LLM can make can increase security. For instance, if you don't enable the delete set of tools, you don't have to worry about the LLM "helping" you by deleting FusionAuth configuration.
  • Lock down your API key. This allows fine-grained control beyond the tool choice. The MCP server communicates to FusionAuth using an API key. You can limit that API key to only allow it to act on tenants, applications and users by choosing the appropriate set of permissions.

Troubleshooting

Verify your API key has the correct permissions for the operation the MCP client is taking.

Check your MCP client logs; these vary by MCP client and platform. For example, $HOME/Library/Logs/Claude/mcp-server-fusionauth-api-server.log is the location for Claude Desktop on macOS. Consult your client's documentation for the precise location.

Use the modelcontextprotocol inspector to help determine if the issue is with the MCP server or with your MCP client: npx @modelcontextprotocol/inspector. If you want to change the USE_TOOLS variable, you cannot dynamically change it and must pass it on the command line. npx @modelcontextprotocol/inspector npx @fusionauth/mcp-api -e API_KEY_APIKEYAUTH=... -e USE_TOOLS=create

Feedback

The FusionAuth team is actively seeking feedback on this preview release.

View the MCP server code or file bug reports by visiting the GitHub repository.

To give feedback about features or share useful prompts, please post in the FusionAuth community forum