Langchain MCP Adapters

integrations/langchain/pyproject.toml

@@ -17,6 +17,9 @@ dependencies = [
     "unitycatalog-langchain[databricks]>=0.3.0",
     "databricks-sdk>=0.65.0",
     "openai>=1.99.9",
+    "langchain-mcp-adapters>=0.1.13",
+    "databricks_mcp>=0.4.0"
+
 ]
 
 [project.optional-dependencies]

integrations/langchain/src/databricks_langchain/__init__.py

@@ -20,6 +20,11 @@
 from databricks_langchain.chat_models import ChatDatabricks
 from databricks_langchain.embeddings import DatabricksEmbeddings
 from databricks_langchain.genie import GenieAgent
+from databricks_langchain.multi_server_mcp_client import (
+    DatabricksMCPServer,
+    DatabricksMultiServerMCPClient,
+    MCPServer,
+)
 from databricks_langchain.vector_search_retriever_tool import VectorSearchRetrieverTool
 from databricks_langchain.vectorstores import DatabricksVectorSearch
 
@@ -34,4 +39,7 @@
     "UnityCatalogTool",
     "DatabricksFunctionClient",
     "set_uc_function_client",
+    "DatabricksMultiServerMCPClient",
+    "DatabricksMCPServer",
+    "MCPServer",
 ]

integrations/langchain/src/databricks_langchain/multi_server_mcp_client.py

@@ -0,0 +1,223 @@
+from typing import Any, Callable, List, Union
+
+from databricks.sdk import WorkspaceClient
+from databricks_mcp.oauth_provider import DatabricksOAuthClientProvider
+from langchain_mcp_adapters.client import MultiServerMCPClient
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class MCPServer(BaseModel):
+    """
+    Base configuration for an MCP server connection using streamable HTTP transport.
+
+    Accepts any additional keyword arguments which are automatically passed through
+    to LangChain's Connection type, making this forward-compatible with future updates.
+
+    Common optional parameters:
+        - headers: dict[str, str] - Custom HTTP headers
+        - timeout: float - Request timeout in seconds
+        - sse_read_timeout: float - SSE read timeout in seconds
+        - auth: httpx.Auth - Authentication handler
+        - httpx_client_factory: Callable - Custom httpx client factory
+        - terminate_on_close: bool - Terminate connection on close
+        - session_kwargs: dict - Additional session kwargs
+
+    Example:
+        ```python
+        from databricks_langchain import DatabricksMultiServerMCPClient, MCPServer
+
+        # Generic server with custom params - flat API for easy configuration
+        server = MCPServer(
+            name="other-server",
+            url="https://other-server.com/mcp",
+            headers={"X-API-Key": "secret"},
+            timeout=15.0,
+            handle_tool_error="An error occurred. Please try again.",
+        )
+
+        client = DatabricksMultiServerMCPClient([server])
+        tools = await client.get_tools()
+        ```
+    """
+
+    model_config = ConfigDict(arbitrary_types_allowed=True, extra="allow")
+
+    name: str = Field(..., exclude=True, description="Name to identify this server connection")
+    url: str
+    handle_tool_error: Union[bool, str, Callable[[Exception], str], None] = Field(
+        default=None,
+        exclude=True,
+        description=(
+            "How to handle errors raised by tools from this server. Options:\n"
+            "- None/False: Raise the error\n"
+            "- True: Return error message as string\n"
+            "- str: Return this string when errors occur\n"
+            "- Callable: Function that takes error and returns error message string"
+        ),
+    )
+
+    def to_connection_dict(self) -> dict[str, Any]:
+        """
+        Convert to connection dictionary for LangChain MultiServerMCPClient.
+
+        Automatically includes all extra fields passed to the constructor,
+        allowing forward compatibility with new LangChain connection fields.
+        """
+        # Get all model fields including extra fields (name is auto-excluded)
+        data = self.model_dump()
+
+        # Add transport type (hardcoded to streamable_http)
+        data["transport"] = "streamable_http"
+
+        return data
+
+
+class DatabricksMCPServer(MCPServer):
+    """
+    MCP server configuration with Databricks authentication.
+
+    Automatically sets up OAuth authentication using the provided WorkspaceClient.
+    Also accepts any additional connection parameters as keyword arguments.
+
+    Example:
+        ```python
+        from databricks.sdk import WorkspaceClient
+        from databricks_langchain import DatabricksMultiServerMCPClient, DatabricksMCPServer
+
+        # Databricks server with automatic OAuth - just pass params as kwargs!
+        server = DatabricksMCPServer(
+            name="databricks-prod",
+            url="https://your-workspace.databricks.com/mcp",
+            workspace_client=WorkspaceClient(),
+            timeout=30.0,
+            sse_read_timeout=60.0,
+            handle_tool_error=True,  # Return errors as strings instead of raising
+        )
+
+        client = DatabricksMultiServerMCPClient([server])
+        tools = await client.get_tools()
+        ```
+    """
+
+    workspace_client: WorkspaceClient | None = Field(
+        default=None,
+        description="Databricks WorkspaceClient for authentication. If None, will be auto-initialized.",
+        exclude=True,
+    )
+
+    def __init__(self, **data):
+        """Initialize DatabricksServer with auth setup."""
+        super().__init__(**data)
+
+        # Set up Databricks OAuth authentication after initialization
+        if self.workspace_client is None:
+            self.workspace_client = WorkspaceClient()
+
+        # Store the auth provider internally
+        self._auth_provider = DatabricksOAuthClientProvider(self.workspace_client)
+
+    def to_connection_dict(self) -> dict[str, Any]:
+        """
+        Convert to connection dictionary, including Databricks auth.
+        """
+        # Get base connection dict
+        data = super().to_connection_dict()
+
+        # Add Databricks auth provider
+        data["auth"] = self._auth_provider
+
+        return data
+
+
+class DatabricksMultiServerMCPClient(MultiServerMCPClient):
+    """
+    MultiServerMCPClient with simplified configuration for Databricks servers.
+
+    This wrapper provides an ergonomic interface similar to LangChain's API while
+    remaining forward-compatible with future connection parameters.
+
+    Example:
+        ```python
+        from databricks.sdk import WorkspaceClient
+        from databricks_langchain import (
+            DatabricksMultiServerMCPClient,
+            DatabricksMCPServer,
+            MCPServer,
+        )
+
+        client = DatabricksMultiServerMCPClient(
+            [
+                # Databricks server with automatic OAuth - just pass params as kwargs!
+                DatabricksMCPServer(
+                    name="databricks-prod",
+                    url="https://your-workspace.databricks.com/mcp",
+                    workspace_client=WorkspaceClient(),
+                    timeout=30.0,
+                    sse_read_timeout=60.0,
+                    handle_tool_error=True,  # Return errors as strings instead of raising
+                ),
+                # Generic server with custom params - same flat API
+                MCPServer(
+                    name="other-server",
+                    url="https://other-server.com/mcp",
+                    headers={"X-API-Key": "secret"},
+                    timeout=15.0,
+                    handle_tool_error="An error occurred. Please try again.",
+                ),
+            ]
+        )
+
+        tools = await client.get_tools()
+        ```
+    """
+
+    def __init__(self, servers: List[MCPServer], **kwargs):
+        """
+        Initialize the client with a list of server configurations.
+
+        Args:
+            servers: List of MCPServer or DatabricksMCPServer configurations
+            **kwargs: Additional arguments to pass to MultiServerMCPClient
+        """
+        # Store server configs for later use (e.g., handle_tool_errors)
+        self._server_configs = {server.name: server for server in servers}
+
+        # Create connections dict (excluding tool-level params like handle_tool_errors)
+        connections = {server.name: server.to_connection_dict() for server in servers}
+        super().__init__(connections=connections, **kwargs)
+
+    async def get_tools(self, server_name: str | None = None):
+        """
+        Get tools from MCP servers, applying handle_tool_error configuration.
+
+        Args:
+            server_name: Optional server name to get tools from. If None, gets tools from all servers.
+
+        Returns:
+            List of LangChain tools with handle_tool_error configurations applied.
+        """
+        import asyncio
+
+        # Determine which servers to load from
+        server_names = [server_name] if server_name is not None else list(self.connections.keys())
+
+        # Load tools from servers in parallel
+        load_tool_tasks = [
+            asyncio.create_task(
+                super(DatabricksMultiServerMCPClient, self).get_tools(server_name=name)
+            )
+            for name in server_names
+        ]
+        tools_list = await asyncio.gather(*load_tool_tasks)
+
+        # Apply handle_tool_error configurations and collect tools
+        all_tools = []
+        for name, tools in zip(server_names, tools_list, strict=True):
+            if name in self._server_configs:
+                server_config = self._server_configs[name]
+                if server_config.handle_tool_error is not None:
+                    for tool in tools:
+                        tool.handle_tool_error = server_config.handle_tool_error
+            all_tools.extend(tools)
+
+        return all_tools