Mossaka avatar

mcp-gateway

Comprehensive management of MCP Gateway (gh-aw-mcpg) for running GitHub MCP servers. Use when: (1) S

提供方 Mossaka|开源

MCP Gateway Management Skill

This skill provides comprehensive guidance for setting up, managing, and debugging the MCP Gateway (gh-aw-mcpg) for running GitHub MCP servers.

What is MCP Gateway?

The MCP Gateway (gh-aw-mcpg) is a proxy server that:

  • Routes requests to multiple MCP (Model Context Protocol) backend servers
  • Manages MCP server lifecycle (start/stop containers)
  • Provides HTTP/SSE transport for MCP protocol
  • Handles authentication and session management
  • Supports both routed (/mcp/<server>) and unified (/mcp) modes

Quick Start

1. Start the Gateway

python3 scripts/start-gateway.py

This will:

  • Get GitHub token from gh CLI (or use $GITHUB_PERSONAL_ACCESS_TOKEN)
  • Create MCP Gateway configuration
  • Start gateway container on port 80
  • Spawn GitHub MCP server container
  • Display connection information

Key Environment Variables:

  • GATEWAY_VERSION - Gateway version (default: v0.0.59)
  • GATEWAY_PORT - Host port (default: 80)
  • GATEWAY_DOMAIN - Gateway domain (default: localhost)
  • GATEWAY_API_KEY - API authentication key (auto-generated)
  • GITHUB_PERSONAL_ACCESS_TOKEN - GitHub PAT for MCP server

2. Verify Gateway is Running

curl http://127.0.0.1:80/health | jq .

Expected response:

{
  "status": "healthy",
  "specVersion": "1.5.0",
  "gatewayVersion": "v0.0.59",
  "servers": {
    "github": {
      "status": "running",
      "uptime": 0
    }
  }
}

3. Debug Issues

python3 scripts/debug-gateway.py

Shows:

  • Container status
  • Health check results
  • Gateway logs
  • MCP server containers
  • Common configuration issues

4. Stop the Gateway

python3 scripts/stop-gateway.py

Stops and removes the gateway container and any MCP server containers.

Configuration

Gateway Config Format (v0.0.59+)

The gateway requires JSON configuration via stdin:

{
  "mcpServers": {
    "github": {
      "type": "stdio",
      "container": "ghcr.io/github/github-mcp-server:latest",
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxx"
      }
    }
  },
  "gateway": {
    "port": 8000,
    "domain": "localhost",
    "apiKey": "your-api-key"
  }
}

Key Points:

  • type: "stdio" - Containerized stdio-based MCP server
  • container - Docker image for the MCP server
  • env.GITHUB_PERSONAL_ACCESS_TOKEN - Must be explicit value (not empty string)
  • gateway.domain - Required field in v0.0.59+
  • gateway.apiKey - Used for client authentication

Required Environment Variables (Container)

When starting the container, you MUST set:

  • MCP_GATEWAY_PORT=8000 - Internal port (must match config)
  • MCP_GATEWAY_DOMAIN=localhost - Gateway domain (must match config)
  • MCP_GATEWAY_API_KEY=xxx - API key (must match config)
  • GITHUB_PERSONAL_ACCESS_TOKEN=xxx - For token passthrough

Starting the Container

Important: The container must be started with -i flag and config piped via stdin:

cat config.json | docker run -i \
  --name mcpg-gateway \
  -p 80:8000 \
  -v /var/run/docker.sock:/var/run/docker.sock \
  -e "GITHUB_PERSONAL_ACCESS_TOKEN=${TOKEN}" \
  -e "MCP_GATEWAY_PORT=8000" \
  -e "MCP_GATEWAY_DOMAIN=localhost" \
  -e "MCP_GATEWAY_API_KEY=${API_KEY}" \
  ghcr.io/githubnext/gh-aw-mcpg:v0.0.59

Note: Do not use -d (detached) with -i - run in background using & instead:

(cat config.json | docker run -i ...) > /tmp/gateway.log 2>&1 &

Using with Copilot CLI

Setup

python3 scripts/setup-copilot.py

This creates /tmp/mcp-gateway-config.json:

{
  "mcpServers": {
    "github-gateway": {
      "type": "http",
      "url": "http://host.docker.internal:80/mcp/github",
      "headers": {
        "Authorization": "your-api-key"
      },
      "tools": ["*"]
    }
  }
}

Usage (Direct)

npx @github/copilot \
  --disable-builtin-mcps \
  --additional-mcp-config @/tmp/mcp-gateway-config.json \
  --allow-all-tools \
  --prompt "List 3 recent issues from githubnext/gh-aw"

Usage (With AWF)

sudo -E awf \
  --enable-host-access \
  --mount /tmp:/tmp:rw \
  --allow-domains 'host.docker.internal,api.github.com,*.githubusercontent.com' \
  -- npx -y @github/copilot \
    --disable-builtin-mcps \
    --additional-mcp-config @/tmp/mcp-gateway-config.json \
    --allow-all-tools \
    --prompt "Your prompt here"

Key Flags:

  • --enable-host-access - Enables host.docker.internal resolution
  • --mount /tmp:/tmp:rw - Makes MCP config accessible in container
  • --disable-builtin-mcps - Prevents Copilot from spawning its own MCP server
  • --additional-mcp-config - Uses HTTP-based gateway config

Troubleshooting

Gateway Won't Start

Symptom: Container immediately exits or restarts

Check:

  1. Are all required env vars set? (MCP_GATEWAY_PORT, MCP_GATEWAY_DOMAIN, MCP_GATEWAY_API_KEY)
  2. Is config being piped via stdin? (container must use -i flag)
  3. Is config format valid JSON?

Debug:

docker logs mcpg-gateway

Look for:

  • "Required environment variables not set"
  • "Container was not started with -i flag"
  • "Configuration validation error"

GitHub MCP Server Fails to Launch

Symptom: Logs show "failed to register tools from github"

Check:

  1. Is GITHUB_PERSONAL_ACCESS_TOKEN explicitly set in config (not empty string)?
  2. Is the token valid and has required scopes?
  3. Is Docker socket mounted (-v /var/run/docker.sock:/var/run/docker.sock)?

Debug:

docker logs mcpg-gateway 2>&1 | grep -A5 'GITHUB_PERSONAL_ACCESS_TOKEN'

Look for:

  • "✗ WARNING: Env passthrough for GITHUB_PERSONAL_ACCESS_TOKEN requested but NOT FOUND"
  • "❌ MCP Connection Failed"

Fix: Use explicit token value in config instead of variable expansion:

{
  "env": {
    "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_actualtoken123"
  }
}

Copilot Can't Connect to Gateway

Symptom: "fetch failed" or "ECONNREFUSED" errors

Check:

  1. Is gateway running? curl http://127.0.0.1:80/health
  2. Is --enable-host-access flag set (for AWF)?
  3. Is host.docker.internal in allowed domains?
  4. Is MCP config file path correct and mounted?

Debug:

# From within AWF container
curl -v http://host.docker.internal:80/health

"method is invalid during session initialization"

Symptom: Tool calls fail with initialization error

Cause: HTTP sessions are stateless - each request creates a new session

Solution: This is expected behavior for HTTP transport. The gateway handles initialization automatically for each request. If you see persistent errors, check:

  1. Is the Authorization header correct?
  2. Is the MCP Gateway version latest?

Port 80 Already in Use

Symptom: "address already in use" when starting container

Solution: Use different host port:

GATEWAY_PORT=8080 python3 scripts/start-gateway.py

Then update Copilot config to use port 8080.

Architecture

┌─────────────────────────────────────────────────────┐
│                    HOST MACHINE                      │
│                                                      │
│  ┌──────────────────┐                               │
│  │  MCP Gateway     │                               │
│  │  Container       │                               │
│  │  Port 80:8000    │                               │
│  └────────┬─────────┘                               │
│           │ spawns via docker socket                │
│           ▼                                          │
│  ┌──────────────────┐                               │
│  │  GitHub MCP      │                               │
│  │  Server          │                               │
│  │  Container       │                               │
│  └──────────────────┘                               │
│                                                      │
│  Client connects to http://127.0.0.1:80/mcp/github  │
└──────────────────────────────────────────────────────┘

With AWF:

┌─────────────────────────────────────────────────────┐
│                    HOST                              │
│  ┌──────────────┐                                    │
│  │ MCP Gateway  │◄──────────┐                        │
│  │ Port 80      │           │                        │
│  └──────────────┘           │                        │
│                             │                        │
│  ┌──────────────────────────┼──────────────┐         │
│  │        AWF Network       │              │         │
│  │  ┌────────┐    ┌─────────┴──┐          │         │
│  │  │ Agent  │───▶│ Squid      │          │         │
│  │  │        │    │ Proxy      │──────────┘         │
│  │  │Copilot │    │            │ CONNECT to         │
│  │  │  CLI   │    │host.docker.│ host:80            │
│  │  └────────┘    │ internal   │                    │
│  │                └────────────┘                    │
│  └───────────────────────────────────────┘         │
└─────────────────────────────────────────────────────┘

Version Differences

v0.0.59 (Latest)

New Requirements:

  • Config via stdin with -i flag (not file mount)
  • Three required env vars: MCP_GATEWAY_PORT, MCP_GATEWAY_DOMAIN, MCP_GATEWAY_API_KEY
  • gateway.domain field required in config
  • Explicit token values (variable expansion ${VAR} may not work reliably)

v0.0.10 (Older)

Simpler startup:

  • No stdin config required
  • Fewer env vars needed
  • Default configuration works out of the box

Recommendation: Use v0.0.59+ for latest features and security improvements.

Best Practices

  1. Token Management

    • Use gh auth token to get current GitHub token
    • Store API key securely (don't commit to repo)
    • Rotate tokens regularly

  2. Container Lifecycle

    • Run gateway in background with & not -d
    • Redirect logs to file for debugging
    • Use --restart unless-stopped for production

  3. Networking

    • Use port 80 for simplicity (or custom port if needed)
    • Always use host.docker.internal from containers
    • Add to AWF allowed domains

  4. Debugging

    • Always check health endpoint first
    • Use debug script for comprehensive diagnostics
    • Check both gateway and MCP server logs

  5. Security

    • Only use --enable-host-access with trusted workloads
    • Keep API keys secret
    • Use restrictive domain whitelists in AWF

Common Commands

# Start gateway
python3 scripts/start-gateway.py

# Check status
curl http://127.0.0.1:80/health | jq .

# View logs
docker logs mcpg-gateway -f

# Debug issues
python3 scripts/debug-gateway.py

# Setup Copilot
python3 scripts/setup-copilot.py

# Stop gateway
python3 scripts/stop-gateway.py

# Restart gateway
docker restart mcpg-gateway

# Check MCP servers
docker ps | grep mcp

# Clean up everything
docker stop mcpg-gateway && docker rm mcpg-gateway
docker ps -a | grep mcp | awk '{print $1}' | xargs -r docker rm -f

Additional Resources

下载 Skill
mcp-gateway - 适用于 Claude Code 与 Cursor 的 AI 智能体 Skill | Agent Skills