Guides7 min read
What is an MCP gateway?
The Model Context Protocol standardized how agents connect to tools and data sources. It worked — which is exactly why organizations now have MCP servers appearing faster than anyone is tracking them, each holding credentials to something that matters.
The problem MCP created
- Credential sprawl. Every MCP server holds tokens for the system it fronts. Nobody has a complete inventory.
- No permission model. An agent with access to a server generally gets everything that server exposes, not a scoped subset.
- No usage visibility. Which agent called which tool, with what arguments, on whose behalf?
- Prompt injection reach. Tool output flows back into context. A compromised or hostile data source can influence subsequent tool calls.
What an MCP gateway does
- Centralizes discovery. One registry of approved servers instead of per-developer configuration files.
- Scopes permissions.Expose a subset of a server's tools to a given agent or team, rather than all of it.
- Brokers credentials. The gateway holds the secret; agents get scoped, revocable access.
- Logs every call. An audit trail of tool invocations and arguments.
- Enforces limits. Rate and cost caps on expensive or destructive tools.
An MCP gateway governs the tool side; an LLM gateway governs the model side. They are complementary, and both are needed once agents run against real systems. Neural Router preserves tool-call fidelity end to end so an MCP layer behaves identically whether or not a router sits in front of the model.
Related: what agentic coding needs from an API and tool calling.