Introduction

As AI systems evolve, Model Context Protocol (MCP) has emerged as a standard for connecting models with external tools and data. At the heart of this ecosystem, Anthropic introduced MCPB (MCP Bundle) — a packaging format designed to simplify how MCP servers are distributed, installed, and executed.

Think of MCPB as the “app bundle” for MCP servers: it encapsulates everything you need to run an MCP server in a single portable file. This blog explores what MCPB is today, how it works, and where it could go in the future.

What is MCPB?

MCPB stands for MCP Bundle, a single-file distribution format for MCP servers. It packages:

  • Server code (Python, Node.js, etc.)
  • manifest.json (server metadata: tools, resources, config)
  • Dependencies (optional runtime libraries)
  • Assets (icons, screenshots, docs)

This format solves a key pain point: without MCPB, installing an MCP server can be messy (different dependencies, languages, and configs). With MCPB, you just need the .mcpb file — making distribution and installation frictionless.

How MCPB Works

  1. Developer prepares source code
    • Includes manifest.json, dependencies, and optional assets.
  2. Pack into .mcpb file 
    npx -y @anthropic-ai/mcpb pack . my-server.mcpb
  3. Distribute
    • Share via GitHub Release, internal repo, or a future MCP marketplace.
  4. Load into host
    • Claude Desktop, IDE plugins to load the .mcpb file directly.
  5. Run the server
    • The host spins up the MCP server, exposing its tools and resources.

Example manifest.json

Here’s a simplified snippet you might see inside a bundle:

{
  "manifest_version": "0.2",
  "name": "my-mcp",
  "display_name": "My MCP Server",
  "version": "0.3.0",
  "description": "My MCP Server",
  "author": {
    "name": "me",
    "email": "me@example.com"
  },
  "tools": [
    {
      "name": "hello_world",
      "description": "Hello World"
    }
  ],
  "resources": [
    {
      "name": "hello",
      "type": "world"
    }
  ],
  "user_config": {
    "version": {
      "type": "string",
      "title": "Version",
      "description": "Version of the package which is being installed.",
      "default": "0.3.0",
      "required": false
    }
  },
  "license": "Apache 2.0",
}

This manifest defines what the server offers (tools, resources) and how users configure it. Check mcpb examples to see more details.

Why MCPB Matters

  • Portability: One file contains everything.
  • Ease of installation: No dependency hell.
  • Standardization: Hosts know exactly how to load a .mcpb.
  • Ecosystem growth: Easier for developers to publish, easier for users to consume.

Long-Term Vision

If MCP is the operating system for AI-native applications, then MCPB is the application bundle format.
Just as we rely on .jar files in Java, .deb packages in Linux, and .vsix files for VSCode, MCPB has the potential to become the standard way to distribute AI plugins and agents.

It’s not just a packaging format — it’s the foundation for a thriving AI-native app ecosystem.

Conclusion

MCPB is still young, but it already shows promise as the distribution backbone of the MCP ecosystem. By solving installation pain points and paving the way for registries, marketplaces, and AI-driven tooling, MCPB could become the “Docker image” or “npm package” of AI-native development.

For developers: start experimenting with @anthropic-ai/mcpb.
For ecosystem builders: imagine what a world of MCPB-powered apps could look like.
The future of AI observability, orchestration, and interaction may very well be packaged inside a .mcpb.