Solves a real problem if you're running multiple MCP servers in Claude Code or similar environments: instead of dumping every tool definition into context at startup (often 5k to 50k tokens), this gateway loads just three tools that let agents discover and request specific capabilities on demand. You configure downstream MCP servers in one file and access policies in another, then agents call list_servers to see what's available, get_server_tools to fetch definitions, and execute_tool to run them. Supports wildcard patterns for tool matching, per-agent access control with deny-before-allow rules, audit logging, and hot config reloads. Most useful when you have lots of MCP servers configured but each agent or subagent only needs a subset, cutting context usage by 90% or more.
claude mcp add --transport stdio roddutra-agent-mcp-gateway uvx agent-mcp-gateway