Desktop Extensions for One-Click MCP Server Installation in Claude Desktop
Key point
Claude Desktop turned MCP server installation into a single-button action with .mcpb extensions.
Details
Claude Desktop Extensions bundle an MCP server and its dependencies into a single .mcpb (MCP Bundle) file, eliminating the complex manual installation process. Users just need to open the file and click Install — no need to install Node.js or Python, or edit JSON configuration.
On September 11, 2025, the extension name changed from .dxt to .mcpb, but existing .dxt extensions continue to work. The core point is not the name but the simplification of the distribution method.
The extension package operates around a single manifest.json, which contains the name, version, author, server configuration, user configuration, tool and prompt declarations, and compatibility information. Servers support Node.js, Python, and binary forms, and template substitutions like ${__dirname} or ${user_config.api_key} safely inject execution paths and secrets.
Claude Desktop also handles complexity on the installation and operation side.
- Built-in runtime: Claude Desktop bundles Node.js, reducing external dependencies.
- Automatic updates: New versions are updated automatically.
- Secure storage: Sensitive values like API keys are stored in the OS keychain.
- Visibility: Users can check extension descriptions, permissions, and configuration requirements in the UI.
Developers create a manifest with npx @anthropic-ai/mcpb init and generate a .mcpb file with npx @anthropic-ai/mcpb pack. From there, dragging the file into the Claude Desktop settings window opens the installation screen, and users can search and install directly from the curated directory.
Anthropic has open-sourced the Desktop Extension spec and toolchain, along with implementation-related code, types, and schemas for macOS and Windows, hoping the MCPB format will spread beyond Claude to other AI desktop apps.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.