runelay

ExplainerNov 2026

MCP gateway vs MCP server

An MCP server exposes the tools of one app. A gateway sits in front of many servers so every agent connects once.

42xless context at connect
1endpoint for every agent
1 keyper agent, revoke one
6.6 msadded per call

The short version

MCP gatewayMCP server
What it isOne endpoint in front of many serversTools of one app, like GitHub
Agents connectOnceOnce per server, per agent
Context at connectA few tools, 1,045 tokens for 125Every schema, 3,548 for GitHub alone
CredentialsStored once, encryptedSet up in every client
Revoking one agentRevoke its keyRotate every credential
AuditEvery call in one logPer server, if any
Examplesrunelay, Executor, ComposioGitHub, Linear, Notion MCP

What each costs your agent at connect

11 MCP servers, no gateway44,43442x
One MCP server (GitHub)3,5483.4x
runelay1,0451x

Tokens sent before the first prompt · benchmark

How they fit together

A gateway does not replace MCP servers, it is built on them. To your agent it is one MCP server. Behind it, it acts as a client to every upstream server, OpenAPI spec or local process you connect, and keeps their credentials.

  • search_toolsFind a tool across every connected account.
  • call_tool · executeCall one tool, or chain several in sandboxed JavaScript.
  • read_memoryShared notes every agent can read and write.
  • read_skillYour SKILL.md folders, served on demand.

One MCP server is enough when

  • You use one agent and one or two apps.
  • A few thousand extra tokens at connect are fine.
  • Logging in again when you switch tools is no big deal.

You need a gateway when

  • Several agents should see the same accounts.
  • You connect many apps and context is tight.
  • You want one key per agent and one audit log.
  • Memory and skills should travel between agents.

Questions

Is an MCP gateway also an MCP server?

Yes. It speaks MCP to your agents and acts as an MCP client to the servers behind it.

Does a gateway add latency?

Barely. A tool call through runelay takes 6.6 ms median on localhost, on top of the upstream server’s own time.

Can a gateway run local stdio MCP servers?

Yes. runelay runs local MCP servers next to remote ones and OpenAPI specs.

Do I have to change my agents?

No. Point them at one endpoint, /v1/mcp, with their own key. runelay connect does it for you.