Does MCP ClientSession support concurrent tool calls? Clarification on multiplexing and scalability?

[aparnarajan-pf]

Pre-submission Checklist

• I have verified that this discussion would not be more appropriate as an issue in a specific repository
• I have searched existing discussions to avoid duplicates

Discussion Topic

I'm building a high-throughput AI agent platform using the MCP Go SDK where multiple concurrent requests need to call MCP tools. I need clarification on how ClientSession handles concurrent requests.

Current Setup
Go HTTP server handling 10+ concurrent chat requests. (This will grow) Each request may need to call 1-5 MCP tools
Single MCP ClientSession shared across all goroutines

What I Need to Understand

1. Concurrency model: If 10 goroutines call session.CallTool() simultaneously, do all 10 requests get sent immediately over the connection, or do concurrent calls block/queue waiting for previous calls to complete?
2. Physical connections: Does one ClientSession use one physical HTTP connection, or can it utilize multiple connections from the http.Client pool?
3. Scalability: Is a single shared ClientSession designed to handle 100+ concurrent tool calls, or should I implement session pooling? (Create a session pool (5-10 sessions) and distribute requests across them or Create a new session per HTTP request (short-lived sessions))

[findleyr]

(1) Calls are concurrent, and (2) for HTTP transports, this means multiple concurrent HTTP requests. By default, we use http.DefaultClient, but you can configure this with client options for the HTTP transports.

Regarding scalability, I strongly recommend testing for your use case. You may find https://github.com/modelcontextprotocol/go-sdk/tree/31e97ad69edad679c726449c4584353502965cf5/examples/client/loadtest to be a good starting point (though looking at that, each worker has a distinct client session; perhaps we should update it to also make concurrent requests on a single session).