gh-aw-mcpg

gh-aw-mcpg is GitHub’s open-source MCP Gateway: a Go service that gives AI agents running in sandboxed environments controlled access to Model Context Protocol servers. Maintained inside GitHub’s own organization and updated throughout 2026, it has earned more than 180 stars on GitHub.

Overview

The gateway was built for GitHub Agentic Workflows, where it is wired in through the sandbox.mcp configuration so that agents executing inside a sandbox can reach external MCP servers through a single, governed endpoint. Instead of letting every agent talk to every tool server directly, gh-aw-mcpg sits in the middle: it launches or connects to the MCP servers listed in its configuration — local stdio servers packaged as containers, or remote servers — and proxies requests to them in routed mode over HTTP. Configuration arrives as a JSON or TOML file, or over standard input, and is validated against a strict schema at startup, with support for per-server environment variables, mounts, entrypoint arguments and connect and tool timeouts. The gateway ships as a Docker image published to GitHub’s container registry, and writes its activity to log files so operators can audit what agents requested and when.

Key Features

• Single governed endpoint – agents reach every configured MCP server through one gateway instead of direct connections.
• Containerized MCP servers – run stdio tool servers as Docker containers with explicit environment, mounts and entrypoints.
• Schema-validated config – JSON and TOML configuration is checked at startup, catching mistakes before agents run.
• Routed HTTP mode – proxies MCP traffic over HTTP on a local listener or a published container port.
• Built for sandboxes – designed for GitHub Agentic Workflows, with per-agent identity in the gateway configuration.
• Operational logging – gateway activity is written to log files for later auditing.
• MIT licensed Go service – small, fast and easy to deploy anywhere Docker runs.

💳 Pricing

gh-aw-mcpg is free and open source under the MIT license. There is no hosted version and no paid tier — you run the gateway yourself, next to your agents.

Who Is It For

It is for platform and DevOps engineers who run AI agents in sandboxes and need a controlled, auditable bridge between those agents and MCP tool servers — especially teams adopting GitHub Agentic Workflows.