[Feat] MCP - Add support for streamablehttp_client MCP Servers

.circleci/config.yml

@@ -785,7 +785,7 @@ jobs:
             pip install "pytest-asyncio==0.21.1"
             pip install "respx==0.22.0"
             pip install "pydantic==2.10.2"
-            pip install "mcp==1.5.0"
+            pip install "mcp==1.9.3"
       # Run pytest and generate JUnit XML report
       - run:
           name: Run tests
@@ -920,7 +920,7 @@ jobs:
             pip install "respx==0.22.0"
             pip install "hypercorn==0.17.3"
             pip install "pydantic==2.10.2"
-            pip install "mcp==1.5.0"
+            pip install "mcp==1.9.3"
             pip install "requests-mock>=1.12.1"
             pip install "responses==0.25.7"
             pip install "pytest-xdist==3.6.1"

.circleci/requirements.txt

@@ -12,4 +12,4 @@ pydantic==2.10.2
 google-cloud-aiplatform==1.43.0
 fastapi-sso==0.16.0 
 uvloop==0.21.0
-mcp==1.5.0    # for MCP server  
+mcp==1.9.3    # for MCP server  

litellm/proxy/_experimental/mcp_server/mcp_server_manager.py

@@ -1,7 +1,7 @@
 """
 MCP Client Manager
 
-This class is responsible for managing MCP SSE clients.
+This class is responsible for managing MCP clients with support for both SSE and HTTP streamable transports.
 
 This is a Proxy
 """
@@ -25,6 +25,12 @@
     MCPTransport,
     MCPTransportType,
 )
+
+try:
+    from mcp.client.streamable_http import streamablehttp_client
+except ImportError:
+    streamablehttp_client = None  # type: ignore
+
 from litellm.types.mcp_server.mcp_server_manager import MCPInfo, MCPServer
 
 
@@ -169,16 +175,43 @@ async def _get_tools_from_server(self, server: MCPServer) -> List[MCPTool]:
 
                     # Update tool to server mapping
                     for tool in tools_result.tools:
-                        self.tool_name_to_mcp_server_name_mapping[
-                            tool.name
-                        ] = server.name
+                        self.tool_name_to_mcp_server_name_mapping[tool.name] = (
+                            server.name
+                        )
 
                     return tools_result.tools
         elif server.transport == MCPTransport.http:
-            # TODO: implement http transport
-            return []
+            if streamablehttp_client is None:
+                verbose_logger.error(
+                    "streamablehttp_client not available - install mcp with HTTP support"
+                )
+                raise ValueError(
+                    "streamablehttp_client not available - please run `pip install mcp -U`"
+                )
+            verbose_logger.debug(f"Using HTTP streamable transport for {server.url}")
+            async with streamablehttp_client(
+                url=server.url,
+            ) as (read_stream, write_stream, get_session_id):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+
+                    if get_session_id is not None:
+                        session_id = get_session_id()
+                        if session_id:
+                            verbose_logger.debug(f"HTTP session ID: {session_id}")
+
+                    tools_result = await session.list_tools()
+                    verbose_logger.debug(f"Tools from {server.name}: {tools_result}")
+
+                    # Update tool to server mapping
+                    for tool in tools_result.tools:
+                        self.tool_name_to_mcp_server_name_mapping[tool.name] = (
+                            server.name
+                        )
+
+                    return tools_result.tools
         else:
-            # TODO: throw error on transport found or skip
+            verbose_logger.warning(f"Unsupported transport type: {server.transport}")
             return []
 
     def initialize_tool_name_to_mcp_server_name_mapping(self):
@@ -217,8 +250,30 @@ async def call_tool(self, name: str, arguments: Dict[str, Any]):
                     await session.initialize()
                     return await session.call_tool(name, arguments)
         elif mcp_server.transport == MCPTransport.http:
-            # TODO: implement http transport
-            raise NotImplementedError("HTTP transport is not implemented yet")
+            if streamablehttp_client is None:
+                verbose_logger.error(
+                    "streamablehttp_client not available - install mcp with HTTP support"
+                )
+                raise ValueError(
+                    "streamablehttp_client not available - please run `pip install mcp -U`"
+                )
+            verbose_logger.debug(
+                f"Using HTTP streamable transport for tool call: {name}"
+            )
+            async with streamablehttp_client(
+                url=mcp_server.url,
+            ) as (read_stream, write_stream, get_session_id):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+
+                    if get_session_id is not None:
+                        session_id = get_session_id()
+                        if session_id:
+                            verbose_logger.debug(
+                                f"HTTP session ID for tool call: {session_id}"
+                            )
+
+                    return await session.call_tool(name, arguments)
         else:
             return CallToolResult(content=[], isError=True)
 

pyproject.toml

@@ -55,7 +55,7 @@ pynacl = {version = "^1.5.0", optional = true}
 websockets = {version = "^13.1.0", optional = true}
 boto3 = {version = "1.34.34", optional = true}
 redisvl = {version = "^0.4.1", optional = true, markers = "python_version >= '3.9' and python_version < '3.14'"}
-mcp = {version = "1.5.0", optional = true, python = ">=3.10"}
+mcp = {version = "1.9.3", optional = true, python = ">=3.10"}
 litellm-proxy-extras = {version = "0.2.3", optional = true}
 rich = {version = "13.7.1", optional = true}
 litellm-enterprise = {version = "0.1.7", optional = true}

requirements.txt

@@ -15,7 +15,7 @@ mangum==0.17.0 # for aws lambda functions
 pynacl==1.5.0 # for encrypting keys
 google-cloud-aiplatform==1.47.0 # for vertex ai calls
 anthropic[vertex]==0.21.3
-mcp==1.5.0    # for MCP server
+mcp==1.9.3    # for MCP server
 google-generativeai==0.5.0 # for vertex ai calls
 async_generator==1.10.0 # for async ollama calls
 langfuse==2.45.0 # for langfuse self-hosted logging