Allison introduces new features in the GitHub MCP Server, including tool-specific configuration, improved context reduction, enhanced security via Lockdown mode, and migration to the official Go SDK for the Model Context Protocol.

GitHub MCP Server: Tool-Specific Configuration, Security, and Protocol Upgrades

Author: Allison

Overview

The latest GitHub MCP Server update introduces:

Tool-specific configuration via the X-MCP-Tools header
Enhanced security with Lockdown mode and content sanitization
Full migration to the official MCP Go SDK for improved protocol support

Tool-Specific Configuration

Previously, toolsets in the GitHub MCP Server enabled broad groups of tools, but fine-grained customization wasn’t possible. Now, using the new X-MCP-Tools header on the remote (and a corresponding flag for the local server), users can enable only the tools they need. This:

Reduces context window usage
Minimizes unnecessary tool loading
Enables better cost efficiency and performance

Example (Remote):

"X-MCP-Tools": "get_file_contents,pull_request_read"

Example (Local):

--tools=get_file_contents,pull_request_read

Mix and match tool-specific and toolset-specific configuration modes for further customization. Example use cases and further documentation are available in the server configuration docs.

Context Reduction Benefits

Loading just the most-used tools (3-10) in context can deliver a 60-90% reduction in context window usage vs. the default toolsets, leading to:

Faster response times
Lower operational costs
More effective use of AI-powered agents

Migration to the Official MCP Go SDK

Both local and remote GitHub MCP Servers have migrated fully to the official Go SDK, keeping the server up to date with new MCP spec features. This enables benefits like resource auto-completions for owners, repo names, and file paths.

A special thanks to the mark3labs/mcp-go community, which previously powered the MCP Server.

Security: Lockdown Mode & Content Sanitization

Lockdown Mode

For public repositories, the new Lockdown mode filters content from untrusted contributors, exposing only content from those with push access. It is enabled with:

"X-MCP-Lockdown": "true"

Certain tools will return errors or filter content when invoked by untrusted authors, bolstering protection against supply chain risks.

Content Sanitization Against Prompt Injection

To defend against prompt injection attacks, content is now automatically sanitized:

Invisible Unicode character filtering
HTML sanitization (strip unsafe tags/attributes)
Removal of unrendered markdown text in code fences

Example: A malicious markdown block containing instructions embedded in code will be sanitized so only safe content (e.g., actual code) is passed to the AI model.

Resources and Feedback

Explore the documentation and repository for more guidance:

If you encounter issues or have feedback, open an issue in the GitHub MCP Server repository.

These updates help developers, DevOps engineers, and AI builders to work more securely and efficiently with GitHub-integrated AI tools.

This post appeared first on “The GitHub Blog”. Read the entire article here