build-mcpb

Table of Contents

• When to Use MCPB vs. Remote Connector
• Platform Support
• Five-Step Build Process
• Directory Structures
• CLI Command Reference
• Custom Exclusions with .mcpbignore
• Development Workflow
• User Configuration
• Submitting to Anthropic Directory

Build an MCP Bundle (MCPB)

MCP Bundles (.mcpb) are zip archives containing a local MCP server and a manifest.json. They enable single-click installation in Claude Desktop, similar to browser extensions (.crx, .vsix).

When to Use MCPB vs. Remote Connector

Choose MCPB (local server) for:

• Access to systems behind your firewall (JIRA, internal wikis, private databases)
• Direct filesystem access, local tool integration (Docker, IDEs, databases)
• Seamless auth via existing SSO/browser sessions — no token management
• Works offline; no cloud infrastructure or VPN required
• Enterprise: one-click install with bundled runtime, admin-controlled allowlists

Choose a remote connector for:

• Cloud services and public APIs requiring centralized infrastructure
• Distribution across Claude Web, mobile, and desktop
• OAuth flows requiring server-side token management
• Public-facing integrations used by multiple organizations

Platform Support

Claude Desktop runs on macOS (darwin) and Windows (win32). Declare supported platforms in manifest.json's compatibility section. Best practice: test on both platforms even if you develop on one.

Five-Step Build Process

Step 1: Install the MCPB CLI

npm install -g @anthropic-ai/mcpb

Step 2: Create Your MCP Server

Build a standard MCP server using the MCP SDK (@modelcontextprotocol/sdk). The server communicates via stdio transport — this is the only transport used in MCPB bundles.

npm install @modelcontextprotocol/sdk

For a Node.js server (server/index.js):

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";

const server = new McpServer({
  name: "my-extension",
  version: "1.0.0",
});

server.tool("my_tool", "Description of what this tool does", {}, async () => {
  return { content: [{ type: "text", text: "Hello from MCPB!" }] };
});

const transport = new StdioServerTransport();
await server.connect(transport);

For a Python server (server/main.py):

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("my-extension")

@mcp.tool()
def my_tool() -> str:
    """Description of what this tool does"""
    return "Hello from MCPB!"

if __name__ == "__main__":
    mcp.run()

Step 3: Initialize the Manifest

Run in your project directory:

mcpb init

The interactive wizard prompts for:

• Extension name, author, version, description
• Server type (Node.js, Python, or Binary)
• Entry point path
• Tools configuration
• Keywords, license, repository

This creates manifest.json. You can also create it manually — see the mcpb-manifest skill for the full spec.

Step 4: Pack the Bundle

# Pack current directory (outputs <name>.mcpb)
mcpb pack .

# Pack with custom output filename
mcpb pack . my-extension-v1.0.mcpb

mcpb pack automatically:

• Validates manifest.json
• Excludes common dev files (.git, node_modules/.cache, .DS_Store, etc.)
• Creates a compressed .mcpb ZIP file

Step 5: Install and Test in Claude Desktop

Three installation methods:

1. Double-click the .mcpb file
2. Drag-and-drop into the Claude Desktop window
3. File menu → Developer → Extensions → Install Extension → select .mcpb

All methods open an installation UI where users review details, configure settings, and grant permissions.

Directory Structures

Node.js Bundle

my-extension/
├── manifest.json         # Required: metadata and configuration
├── server/
│   └── index.js          # Main entry point
├── node_modules/         # Bundled dependencies (run npm install first)
├── package.json
└── icon.png              # Optional: 512x512 PNG

Python Bundle (with bundled deps)

my-extension/
├── manifest.json
├── server/
│   ├── main.py
│   └── lib/              # Bundled packages (pip install --target server/lib .)
└── icon.png

Python Bundle (with uv — experimental, manifest v0.4)

my-extension/
├── manifest.json         # server.type = "uv"
├── pyproject.toml        # Dependencies declared here
├── src/
│   └── server.py
└── .mcpbignore           # Exclude .venv, __pycache__, etc.

The uv type is experimental (v0.4+). The host manages Python installation — small bundle size (~100 KB vs 5–10 MB for bundled deps).

CLI Command Reference

Command	Description
mcpb init [dir]	Interactive manifest creation wizard
mcpb validate <path>	Validate manifest.json against schema
mcpb pack <dir> [output]	Pack directory into .mcpb file
mcpb sign <file>	Sign .mcpb with X.509 certificate
mcpb verify <file>	Verify signature of signed .mcpb
mcpb info <file>	Show file size and signature details
mcpb unsign <file>	Remove signature (dev/testing)

Validate Before Packing

mcpb validate manifest.json
# or
mcpb validate .

Signing (Optional but Recommended for Distribution)

# Self-signed (development/testing)
mcpb sign my-extension.mcpb --self-signed

# Production (with CA certificate)
mcpb sign my-extension.mcpb \
  --cert production-cert.pem \
  --key production-key.pem \
  --intermediate intermediate-ca.pem

# Verify after signing
mcpb verify my-extension.mcpb

Custom Exclusions with .mcpbignore

Create .mcpbignore to exclude additional files from the bundle:

# .mcpbignore
*.test.js
src/**/*.test.ts
coverage/
docs/
.env*
temp/

Development Workflow

# 1. Develop your MCP server
mkdir my-extension && cd my-extension

# 2. Initialize
mcpb init

# 3. Implement server code

# 4. For Node.js: install and bundle dependencies
npm install

# 5. Validate
mcpb validate .

# 6. Pack
mcpb pack .

# 7. Install in Claude Desktop (double-click the .mcpb file)

# 8. Test tools via Claude Desktop

User Configuration

If your server needs user-provided values (API keys, directories, settings), define user_config in manifest.json. Claude Desktop auto-generates a settings UI. Reference values via ${user_config.KEY} in mcp_config.args or mcp_config.env.

See the mcpb-manifest skill for full user_config schema and examples.

Submitting to Anthropic Directory

To distribute via the Anthropic Connectors Directory, additional requirements apply:

• Mandatory tool annotations for all tools
• Minimum three working examples
• Testing credentials (if applicable)

See Local MCP Server Submission Guide.