MCP Gateway

A service orchestration layer for the Model Context Protocol. Manages multiple MCP services through a single unified interface — spawn, supervise, and route tool calls across all your services.

Key Feature: Lazy Activate

By default, services activate in lazy mode — the process spawns and tool list is fetched, but zero tool schemas are injected into the LLM context. This means you can have hundreds of tools across dozens of services with no token overhead.

activate({name: "google-workspace"})     → 281 tools ready, 0 context tokens
tools({service: "google-workspace"})     → browse available tools
call({service: "...", tool: "...", args}) → call any tool directly

Features

• Lazy activate — spawn services with 0 context tokens, browse with tools(), call with call()
• Service orchestration — spawn, stop, restart child MCP services
• On-demand tool browsing — tools({service, filter}) returns tool descriptions as conversation text (not system prompt)
• Direct routing — call() routes to child connections without needing schema registration
• Environment variable expansion — use $VAR or ${VAR} in config for secrets
• Dynamic management — add/remove services at runtime without restarting
• Health monitoring — built-in ping/health checks
• Any command — supports bun, node, npx, ssh, or any executable

Quick Start

# Install
bun install

# Create your config from the example
cp gateway.config.example.json gateway.config.json

# Edit gateway.config.json with your services
# Then start the gateway
bun run start

Configuration

gateway.config.json defines your services:

{
  "services": [
    {
      "name": "my-service",
      "command": "bun",
      "args": ["run", "/path/to/service/index.ts"],
      "env": { "API_KEY": "$MY_API_KEY" },
      "autoActivate": false
    }
  ]
}

Field	Required	Default	Description
name	yes	—	Unique service identifier
command	yes	—	Executable to run (bun, node, npx, ssh, ...)
args	no	[]	Command arguments
env	no	{}	Environment variables (supports $VAR expansion)
autoActivate	no	false	Start automatically on gateway launch
keepAlive	no	false	Always-on: respawn with exponential backoff on crash; never GC'd when idle
groups	no	—	Named tool subsets for full-mode activation, e.g. {"gmail": ["send", "search"]}

Environment Variables

Config values support $VAR and ${VAR} syntax, resolved from process.env at load time:

{
  "env": { "API_KEY": "$MY_SECRET_KEY" },
  "args": ["--config", "${HOME}/.config/my-service.json"]
}

This keeps secrets out of your config file. Pass them through your MCP client config (see below).

Built-in Tools

Once running, the gateway exposes these management tools:

Tool	Description
services	List all services with status, mode (lazy/full), tool count, uptime
activate	Start a service. Default lazy (0 context tokens). Set lazy=false for full schema registration
tools	List available tools for an active service. Supports filter for keyword search
deactivate	Stop a service and clean up
restart	Kill and respawn a service (preserves lazy/full mode)
health	Ping all active services
add	Register a new service dynamically (persists to config)
remove	Remove a service from config
call	Call any tool on any active service

Workflow

1. activate({name: "my-service"})                        → spawn process, 0 tokens
2. tools({service: "my-service"})                        → see all tools
3. tools({service: "my-service", filter: "search"})      → filter by keyword
4. call({service: "my-service", tool: "...", args: {}})   → call a tool
5. deactivate({name: "my-service"})                      → stop when done

Full Mode (optional)

If you want tool schemas injected into the LLM context (traditional MCP behavior):

activate({name: "my-service", lazy: false})              → register all schemas
activate({name: "my-service", lazy: false, groups: ["gmail", "drive"]})  → register specific groups

HTTP Daemon Mode (always-on, shared)

Set GATEWAY_HTTP_PORT to run the gateway as a long-lived Streamable HTTP daemon instead of per-client stdio:

GATEWAY_HTTP_PORT=8770 bun run start

• Endpoint: http://127.0.0.1:8770/mcp (Streamable HTTP, Mcp-Session-Id per client)
• Health: GET /health returns JSON with session count and per-service status/pid
• Shared processes, isolated views: child services are spawned once and shared by every connected client; each session keeps its own tool registry, so one agent's full-mode activation never leaks into another's context
• Session GC: sessions idle longer than the TTL are swept automatically (clients often exit without sending DELETE). Tune with GATEWAY_SESSION_TTL_MS (default 60 min) and GATEWAY_SESSION_SWEEP_MS (default 5 min)
• Safe boot: GATEWAY_NO_AUTOACTIVATE=1 skips autoActivate services (e.g. when stateful singletons are already running elsewhere)

A systemd unit is provided in deploy/mcp-gateway.service.

Point Claude Code (or any MCP client) at the daemon:

{
  "mcpServers": {
    "gateway": { "type": "http", "url": "http://127.0.0.1:8770/mcp" }
  }
}

Using with Claude Code

Add to your Claude Code MCP config (~/.claude/config.json or project settings). API keys go in the env block — they're passed to the gateway process and expanded in gateway.config.json via $VAR syntax:

{
  "mcpServers": {
    "gateway": {
      "command": "bun",
      "args": ["run", "/path/to/mcp-gateway/src/index.ts"],
      "env": {
        "GEMINI_API_KEY": "your-gemini-key",
        "EXA_API_KEY": "your-exa-key"
      }
    }
  }
}

License

MIT