MCP clients and multi-server routing
What the host does when several servers are connected at once.
4 min read · Lesson 8 of 10 in this domain
A host that connects to several MCP servers creates one client per server, each holding its own dedicated connection. That one-to-one mapping is what makes multi-server setups tractable: each connection has its own lifecycle, its own capability negotiation and its own failure domain, so a server going down takes out its client and nothing else. The host is then responsible for aggregation — presenting tools from several servers to the model, and routing a chosen call back to the right client. Since two servers can legitimately expose similarly named tools, disambiguation belongs to the host, and the practical lever is the same as always: descriptions clear enough that selection is unambiguous.
- One client per server, each with a dedicated connection — there is no multiplexing.
- Each connection negotiates capabilities independently, so servers can offer different subsets of primitives.
- The host aggregates tools from all connected servers and routes an invocation back to the originating client.
- Name collisions across servers are a real risk; disambiguation is the host's job and descriptions are what make it possible.
- A failing server degrades only its own client, which is an argument for separate servers over one monolith.
- As the aggregate surface grows, selection accuracy degrades — the same reason dynamic tool discovery exists.
"One client multiplexes every server" is the persistent misreading. It is one dedicated client per server.
Three MCP servers are connected. How many clients does the host create?
Connections are strictly one-to-one, so each server gets its own dedicated client. That is what gives each connection an independent lifecycle and failure domain.
Two connected servers both expose a tool called 'search'. Who resolves the ambiguity?
Nothing in the protocol prevents two servers using the same tool name, so the host must disambiguate — and descriptions are the signal the model uses to choose correctly.
Practise this domain with 18%%-weighted questions in the study app.
Open in study appSource: MCP — Architecture overview · Independent study aid, not affiliated with or endorsed by Anthropic.