MCPToolset raises MCPConnectionError in init if authentication fails

[julian-risch]

When MCPToolset is initialized, we try to make a connection to the MCP server and raise an MCPConnectionError if no connection can be established.

mcp_toolset = MCPToolset(
    server_info=server_info,
    tool_names=["get_current_time"]  
)

This is problematic when a pipeline gets validated, for example to validate a pipeline template that uses an MCPToolset with a GitHub MCP server requiring authentication:

        tools:
          data:
            server_info:
              type: haystack_integrations.tools.mcp.mcp_tool.StreamableHttpServerInfo
              url: https://api.githubcopilot.com/mcp/
              token:
                env_vars:
                - GITHUB_PAT
                strict: false
                type: env_var
              timeout: 900
            tool_names: [get_issue, list_pull_requests]
          type: haystack_integrations.tools.mcp.MCPToolset

full example here with the error here.
The problem is that at the time of pipeline validation, the authentication token to connect to the MCP server, is not set yet. We need to enable pipeline validation without raising an error connecting to the MCP server. The pipeline validation calls the init of MCPToolset.
Other Haystack components have a warmup method. Maybe that could be an option here to reduced the init? Or have a parameter in the init that let's users specify if a connection should be established at initialization time.

[vblagoje]

Ok so the crux of the problem is eager connection attempts during initialization which breaks pipeline validation. The solution: lazy initialization e.g. defer the connection attempts to warmup?

[vblagoje]

Since we don't have warmup in Tool perhaps we can have eager_connect flag in init - if false we connect on first Tool invoke, wdyt?

[julian-risch]

A flag in tool init or toolset init sounds good to me.

[vblagoje]

init parameter alone can't help because MCPTool(set) needs to connect to MCP server and fetch the tool schema. The plan was to connect on first tool invocation but the tool will never get invoked because it needs to be registered with LLM/ToolInvoker. So the conclusion that @julian-risch @sjrl and @anakin87 and I made is to explore the idea of warmup method in Tool(set) that will be called when Tool(set) is passed to LLM/ToolInvoker.