MCP Client Connection Handling on A2A Server

[signoredems]

My implementation includes a Strands A2A server that utilizes tools from a remote MCP server that uses SSE (FWIW, this is the Atlassian MCP server but I'm not sure that detail matters). The A2A Server must use multiple MCP tools when a client is interacting with it, utilizing the results from one tool call to make subsequent ones, but likely has long periods of time in between client connections. What is the correct way to manage connections in this scenario? I can go into how I handle this currently but I'd rather leave it open-ended in the hopes that there is a straightforward way of implementing this that I've totally missed.

Note: I hopeful that I could just use the improved MCP DX feature but I don't see that in the roadmap.

[signoredems]

How I handle this currently is by creating a custom class that inherits from A2AServer. I went this route because I can't call mcp_client.list_tools_sync() to populate the agent's list of tools (which in turn is used to list the skills on the A2A agent card) when the agent is initialized since the MCP server requires OAuth consent. I created each MCPAgentTool manually with a spec that matches the the corresponding tool on the MCP server and add these as my agent tools. I don't see a way around this or something similar.

The custom class has these callbacks for the tool invocation events:

• BeforeToolInvocationEvent: Since the agent only has access to MCPAgentTools, I know a connection to the MCP server will be required so I call self.MCPClient.start().
• AfterToolInvocationEvent: I call self.MCPClient.stop().

This causes an unrecoverable error when the agent tries to run several tools in parallel (#664). I avoid this by using the SequentialToolExecutor but this isn't ideal since the tools should be able to run concurrently in certain scenarios.

One alternative I tried was to create a ContextManagedMCPAgentTool class that inherits from MCPAgentTool, overrides the stream() function, and wraps the call to self.MCPClient.call_tool_async() in a context manager. This seemed like a better-but-not-great implementation (ideally the tool calls would share one connection to the MCP server rather than creating many) but has the same issue when the agent tries to make parallel tool calls.

What is the recommended (or at least a better) way of doing this? The mcp_client tool from the Strands Agents Tools package looks promising but, as far as I can tell, I'd need to manually tell it to connect to the MCP Server rather than just giving it a URL and letting the agent decide when to do this which leaves me essentially where I already am.