An MCP gateway is a single Model Context Protocol endpoint that sits between an AI agent and many tools, so the agent loads one small tool surface instead of every server's full schema.
Without a gateway, each MCP server is its own connection: its own URL or command, its own credentials, and its own tool definitions loaded into the model's context at connect time. Ten servers means ten setups and ten schema payloads. A gateway collapses that into one connection and one credential, and decides which underlying tool runs for a given call.
Why it matters
Every tool definition an agent loads is context it pays for on every turn. A ToolRouter measurement on 3 September 2026 put its own hosted surface at 47 tool definitions and 9,286 tokens, against 257,047 tokens for exposing the same catalog as one MCP tool per skill — a 27.7x difference. The gateway keeps that surface fixed: adding tools to the catalog adds rows to a search result, not schemas to the connection.
The second reason is operational. One endpoint means one place to hold credentials, one bill, one audit trail, and one thing to allow through a corporate policy — instead of a new server to review each time someone wants a new capability.
A gateway is not free of trade-offs. You are trusting one party with routing and credentials, and a tool the gateway does not carry is a tool you cannot call through it.
How ToolRouter does it
ToolRouter exposes a fixed set of gateway tools over one endpoint at https://api.toolrouter.com/mcp. Two of them do catalog work: discover searches the catalog by intent, and use_tool runs a named skill. The rest are account, credential, file, job and memory operations the agent uses directly.
Related
- MCP gateway for companies — the product page, including team billing and audit
- What is a tool router?
- What is a remote MCP server?
- How much context do MCP tool schemas cost? — the measurement behind the numbers above
- ToolRouter glossary