Distribute your MCP server

Once you’ve built an MCP server, you need to decide how to distribute it to users. The four most common ways to expose an MCP server are:

• As an open-source project
• As an npm package
• As an installable MCPB file
• As a remote server

Which of these methods you select depends on two factors:

• Your users: Are they technical developers or non-technical business users?
• Your resources: Can you maintain and pay for a hosted service?

If your users are technical, distribute your MCP server as an open-source project that developers can build themselves or as an npm package for easier installation.

Open-source projects require you to provide build instructions and are best suited for technical users.

MCP servers built using JavaScript or TypeScript can be packaged with npm. For users, this is more convenient than building a project locally but still requires writing some configuration.

If your users are non-technical, and you can’t maintain a hosted server, use MCPB packaging. MCPB (previously called DXT) creates a desktop installer that users can install in Claude Desktop with minimal steps.

If you have the financial and developer resources for hosting, deploy your MCP server remotely. Although remote servers cost more, the minimal configuration for users and easy integration for developers building automated systems means that they provide the best user experience.

Choosing the right distribution method is critical to the successful adoption of your MCP server. This guide explores when best to use each method and demonstrates how to implement it using an example Echo MCP server.

Prerequisites

To follow this tutorial, you need:

• A working understanding of MCP transports and how the MCP server receives inputs and sends outputs
• Node.js 22 or later
• A clone of the simple MCP Echo server, which demonstrates all distribution methods without complex dependencies

Install the MCP Echo server as follows:

Distribute your MCP server as an open-source project

The simplest option is to release your MCP server as an open-source project and provide build and installation instructions in a README. This works when:

• Your users are technical developers comfortable with command-line tools
• You don’t have resources for maintaining a remote server
• You’re starting out and would benefit from community input and contributions

Once the project matures, consider deploying your MCP server as an npm package or MCPB extension so that you can reach non-technical users.

Upload your MCP server to an open-source repository

First, upload your project to an artifact repository platform like GitHub .

Then, create a README that provides your users with build steps and configuration details for their AI tools, similar to the instructions in the example Echo MCP server README below.

Note that the mcp-echo-input-server project uses plain JavaScript, so no build step is needed. If your project uses TypeScript, add a build step before installing the dependencies. For example:

Build the project with the following command:

Example README for the Echo MCP server

Install all the project dependencies:

The entry point file is index.js.

Add the following MCP server configuration to Claude Desktop, via Settings -> Developer -> Edit Config, after replacing <USERNAME> and <PATH-TO-PROJECT> with your actual username and the absolute path to the project:

Reload Claude Desktop, open a new conversation and enter the following prompt:

You will receive an output similar to the following:

Distribute your MCP server as an npm package

If you built your MCP project with JavaScript or TypeScript, you can publish your MCP server as an npm package. This works when:

• Your users are technical developers who can handle npm installations
• You don’t have the resources to host and maintain a remote server

If you’re creating an MCP server that exposes API capabilities and you have an OpenAPI document, you can use Speakeasy to generate a TypeScript MCP server ready for deployment and distribution.

Prepare your MCP server for packaging

Install the project dependencies:

Next, make sure the package.json file is ready for packaging. Add a name, description, author name, author email address, keywords, entry point (main), and the CLI command bin to run the project:

Check that your package name is available

The npm registry requires unique package names. You need to check whether your package name is available before publishing.

First, ensure that you’re logged in:

Then run the following command with the name of your project in the package.json file.

If it returns a 404 error, proceed with the rest of the commands. Otherwise, try other combinations until you find an unused name, then update the name in the package.json file, and proceed.

Publish your MCP server as an npm package

Package and publish your npm package:

When you publish the package, include documentation that provides users with the instructions for installing the MCP server in Claude Desktop and the configuration code they need to add to their claude_desktop_config.json files.

Example documentation for the Echo MCP server

Add the following configuration to your claude_desktop_config.json file:

Reload Claude Desktop. You should now see the MCP server in the chat. Enable it using the toggle button next to echo-server.

Enter the following prompt to test the MCP server functionality:

Claude should return a similar response to the following:

Automate releases

To make release and distribution easier, consider using CI/CD tools, such as this GitHub Action , to automate the npm release process.

Distribute your MCP server as an MCPB file

MCP Bundles (MCPB) , formerly known as Desktop Extensions (DXT), is a zip archive extension developed by Anthropic to make MCP server installation easier with one click. An MCPB package contains the local MCP server and a manifest.json file describing its capabilities.

Note: If you previously used DXT, consult the migration instructions in the MCPB repository README .

MCPB is useful when:

• You’re targeting non-technical users who need your MCP server but shouldn’t have to handle npm installations or dependency management: For example, business users, marketing teams, or domain experts
• You want distribution without infrastructure costs: Local installation means no hosting, scaling, or maintenance overhead on your end, while still providing a professional installation experience
• You’re building cross-platform tools: MCPB packages work across operating systems and can bundle servers written in any language (such as Python, Node.js, or Go), as long as they implement the MCP protocol

Package your MCP server as an MCPB file

First, install the project dependencies:

Then install the @anthropic-ai/mcpb package, which contains a CLI that helps with the creation of both the manifest.json and the final .mcpb file:

Although the MCPB documentation suggests installing @anthropic-ai/mcpb globally, you can install it as a dev dependency in your projects to maintain consistent versions.

Create a manifest.json file by running the following interactive command and answering the configuration prompts that it returns:

When prompted, enter values for fields such as the author name, package name, and description, and add the following user-configurable options, which you’ll use for the USER_NAME environment variable:

At the time of writing, a bug prevents Claude from recognizing the manifest_version key in the manifest.json file. Resolve this issue by replacing manifest_version with dxt_version in the generated manifest.json file:

In the same manifest.json file, edit the USER_NAME variable to point to the userName user-configurable option:

Next, create the distributable MCP file by running the pack command:

Test the MCPB file

Find the generated mcp-echo-input-server.mcpb file in the project directory. Let’s open the file with Claude and install the MCP server.

Navigate to the Claude Desktop Settings, open the Extensions screen, and click the Advanced settings button.

In the advanced settings for All extensions, scroll to the bottom of the page and click the Install Extension… button. Browse for and select the mcp-echo-input-server.mcpb file.

When Claude displays a modal with the extension description, confirm the installation by clicking Install.

In the Configure dialog, enter a username value, such as your name, to be used as an environment variable for the project.

Save the changes and enable the extension using the toggle button.

Open a new conversation in Claude, ensuring that the MCP server is activated.

Test the MCP server by prompting Claude to use it:

You should receive a similar output to the following:

Note: You may encounter Claude alerts about typing issues. Ignore them by clicking the X button or prevent them by typing your projects appropriately.

Upload the MCPB file

You now have a working .mcpb file. Upload it to an artifact repository so that users can download and install your MCP server.

You can also add it to your CI/CD pipeline using this GitHub Action , which automatically builds and publishes your .mcpb file to the GitHub repository when there is a new release.

Distribute your MCP server remotely

If you have the resources to maintain a remote server, distributing your MCP server remotely offers the best solution from both technical and business perspectives.

Benefits of remote distribution

Remote distribution allows you to do the following:

• Monitor, log, and track tool calls, authentication, and user activity.

• Implement business models for monetizing your MCP server.

• Maintain control over data in sensitive fields.

• Simplify installation for Claude Desktop users who can add remote servers via Claude Connectors.

  

  If your server requires authentication and your users are technical, provide them with the necessary documentation for entering OAuth configurations. If your users aren’t technical, use the Connect button to redirect them to an OAuth authorization page that grants them access to your API or services.

• Provide integration options for developers building automated workflows with the Anthropic SDK and OpenAI Agents SDK.

Build the server

If you’re exposing your API capabilities through MCP tools, you have two options:

• Build the infrastructure yourself and handle architecture and maintenance.
• Use Gram, the platform we developed for creating, hosting, and distributing MCP servers with just an OpenAPI document.

Choose your hosting platform

You can use traditional cloud servers, such as AWS, GCP, and Azure, to deploy your MCP server remotely. However, this option requires you to handle server setup, configure SSL certificates, and manage server configuration yourself.

If you’ve created your server using Gram , we recommend hosting it there for a more streamlined experience.

Other MCP server creation platforms have similar hosting and deployment solutions, such as FastMCP Cloud .

Set up OAuth authentication

Implement OAuth 2.1 to secure your remote MCP server:

• For Gram-hosted servers, follow the guide to OAuth implementation with Gram .
• For self-hosted servers, follow the guide to general OAuth implementation for MCP .

Distribute to users

Based on your security requirements, decide whether your server will be public or private:

• Private servers work best for APIs used by enterprise teams, such as for internal marketing tools. You handle user configuration variables on the server side and tightly control access.
• Public servers work best for APIs requiring broader distribution, such as for SaaS, but they require solid OAuth authentication. Although users set their own configuration variables, secure authorization flows are still necessary.

Final thoughts

In this guide, we’ve explored how to distribute your MCP server with different distribution options.

MCP server distribution strategies depend on target users and available resources. Follow the recommendations below according to your use case:

• For local installations, provide multiple distribution options like MCPB, local build, and npm packages. This gives developers alternatives if they encounter issues with any single method. MCPB is well-maintained, so non-technical users will rarely have problems.
• In your documentation, provide configuration examples for popular AI agents like Claude, Cline, Cursor, and VS Code.
• Publish your MCP server on the official Anthropic MCP registry. If you’re using Gram, consult our guide to referencing your MCP server in the registry.