fix: only load mcp when enabled in tool_group (#2621)

# What does this PR do?

The agent code is currently importing MCP modules even when MCP isn’t
enabled. Do we consider this worth fixing, or are we treating MCP as a
first-class dependency? I believe we should treat it as such.

If everyone agrees, let’s go ahead and close this.

Note: The current setup breaks if someone builds a distro without
including MCP in tool_group but still serves the agent API.

Also, we should bump the MCP version to support streamable responses, as
SSE is being deprecated.

Signed-off-by: Sébastien Han <seb@redhat.com>

llama_stack/providers/inline/agents/meta_reference/openai_responses.py

@@ -74,7 +74,6 @@ from llama_stack.log import get_logger
 from llama_stack.models.llama.datatypes import ToolDefinition, ToolParamDefinition
 from llama_stack.providers.utils.inference.openai_compat import convert_tooldef_to_openai_tool
 from llama_stack.providers.utils.responses.responses_store import ResponsesStore
-from llama_stack.providers.utils.tools.mcp import invoke_mcp_tool, list_mcp_tools
 
 logger = get_logger(name=__name__, category="openai_responses")
 
@@ -627,6 +626,8 @@ class OpenAIResponsesImpl:
                     raise ValueError(f"Tool {tool_name} not found")
                 chat_tools.append(make_openai_tool(tool_name, tool))
             elif input_tool.type == "mcp":
+                from llama_stack.providers.utils.tools.mcp import list_mcp_tools
+
                 always_allowed = None
                 never_allowed = None
                 if input_tool.allowed_tools:
@@ -760,7 +761,9 @@ class OpenAIResponsesImpl:
         error_exc = None
         result = None
         try:
-            if function.name in ctx.mcp_tool_to_server:
+            if ctx.mcp_tool_to_server and function.name in ctx.mcp_tool_to_server:
+                from llama_stack.providers.utils.tools.mcp import invoke_mcp_tool
+
                 mcp_tool = ctx.mcp_tool_to_server[function.name]
                 result = await invoke_mcp_tool(
                     endpoint=mcp_tool.server_url,

llama_stack/providers/registry/agents.py

@@ -23,7 +23,7 @@ def available_providers() -> list[ProviderSpec]:
                 "pillow",
                 "pandas",
                 "scikit-learn",
-                "mcp",
+                "mcp>=1.8.1",
             ]
             + kvstore_dependencies(),  # TODO make this dynamic based on the kvstore config
             module="llama_stack.providers.inline.agents.meta_reference",

llama_stack/providers/registry/tool_runtime.py

@@ -85,7 +85,7 @@ def available_providers() -> list[ProviderSpec]:
                 adapter_type="model-context-protocol",
                 module="llama_stack.providers.remote.tool_runtime.model_context_protocol",
                 config_class="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderConfig",
-                pip_packages=["mcp"],
+                pip_packages=["mcp>=1.8.1"],
                 provider_data_validator="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderDataValidator",
                 description="Model Context Protocol (MCP) tool for standardized tool calling and context management.",
             ),