client

Class info

Classes

Name	Children	Inherits
MCPClient llmling_agent.mcp_server.client FastMCP-based client for communicating with MCP servers.		object

🛈 DocStrings

FastMCP-based client implementation for LLMling agent.

This module provides a client for communicating with MCP servers using FastMCP. It includes support for contextual progress handlers that extend FastMCP's standard progress callbacks with tool execution context (tool name, call ID, and input).

The key innovation is the signature injection system that allows MCP tools to work seamlessly with PydanticAI's RunContext while providing rich progress information.

MCPClient

FastMCP-based client for communicating with MCP servers.

Source code in src/llmling_agent/mcp_server/client.py

class MCPClient:
    """FastMCP-based client for communicating with MCP servers."""

    def __init__(
        self,
        config: MCPServerConfig,
        elicitation_callback: ElicitationHandler | None = None,
        sampling_callback: ClientSamplingHandler | None = None,
        progress_handler: ContextualProgressHandler | None = None,
        message_handler: MessageHandlerT | MessageHandler | None = None,
        accessible_roots: list[str] | None = None,
    ):
        self._elicitation_callback = elicitation_callback
        self.config = config
        self._sampling_callback = sampling_callback
        self._progress_handler = MultiEventHandler[ContextualProgressHandler](
            handlers=[progress_handler] if progress_handler else []
        )
        # Store message handler or mark for lazy creation
        self._message_handler = message_handler
        self._accessible_roots = accessible_roots or []
        self._client = self._get_client(self.config)
        self._available_tools: list[MCPTool] = []
        self._connected = False

    async def __aenter__(self) -> Self:
        """Enter context manager."""
        try:
            # First attempt with configured auth
            await self._client.__aenter__()

        except Exception as first_error:
            # OAuth fallback for HTTP/SSE if not already using OAuth
            if _should_try_oauth_fallback(self.config):
                try:
                    with contextlib.suppress(Exception):
                        await self._client.__aexit__(None, None, None)
                    self._client = self._get_client(self.config, force_oauth=True)
                    await self._client.__aenter__()
                    logger.info("Connected with OAuth fallback")
                except Exception:  # noqa: BLE001
                    raise first_error from None
            else:
                raise

        self._connected = True
        await self._refresh_tools()
        return self

    async def __aexit__(
        self,
        exc_type: type[BaseException] | None,
        exc_val: BaseException | None,
        exc_tb: TracebackType | None,
    ):
        """Exit context manager and cleanup."""
        try:
            await self._client.__aexit__(None, None, None)
        except Exception as e:  # noqa: BLE001
            logger.warning("Error during FastMCP client cleanup: %s", e)
        finally:
            self._connected = False
            self._available_tools = []

    async def _log_handler(self, message: LogMessage) -> None:
        """Handle server log messages."""
        level = LEVEL_MAP.get(message.level.lower(), logging.INFO)
        logger.log(level, "MCP Server: %s", message.data)

    async def _elicitation_handler_impl(
        self,
        message: str,
        response_type: type,
        params: ElicitRequestParams,
        context: RequestContext,
    ) -> ElicitResult[dict[str, Any]] | dict[str, Any] | None:
        """Handle elicitation requests from server."""
        if not self._elicitation_callback:
            return None

        try:
            # Direct FastMCP callback - no compatibility layer
            return await self._elicitation_callback(
                message, response_type, params, context
            )
        except Exception:
            logger.exception("Elicitation handler failed")
            from fastmcp.client.elicitation import ElicitResult

            return ElicitResult(action="cancel")

    async def _sampling_handler_impl(
        self,
        messages: list[SamplingMessage],
        params: CreateMessageRequestParams,
        context: RequestContext,
    ) -> str:
        """Handle sampling requests from server."""
        if not self._sampling_callback:
            return "Sampling not supported"

        try:
            result = self._sampling_callback(messages, params, context)
            if inspect.iscoroutine(result):
                result = await result
            return str(result)
        except Exception as e:
            logger.exception("Sampling handler failed")
            return f"Sampling failed: {e}"

    def _get_client(self, config: MCPServerConfig, force_oauth: bool = False):
        """Create FastMCP client based on config."""
        import fastmcp
        from fastmcp.client import SSETransport, StreamableHttpTransport
        from fastmcp.client.transports import StdioTransport

        from llmling_agent_config.mcp_server import (
            SSEMCPServerConfig,
            StdioMCPServerConfig,
            StreamableHTTPMCPServerConfig,
        )

        transport: ClientTransport
        # Create transport based on config type
        match config:
            case StdioMCPServerConfig(command=command, args=args):
                env = config.get_env_vars()
                transport = StdioTransport(command=command, args=args, env=env)
                oauth = False
                if force_oauth:
                    msg = "OAuth is not supported for StdioMCPServerConfig"
                    raise ValueError(msg)

            case SSEMCPServerConfig(url=url, headers=headers, auth=auth):
                transport = SSETransport(url=url, headers=headers)
                oauth = auth.oauth

            case StreamableHTTPMCPServerConfig(url=url, headers=headers, auth=auth):
                transport = StreamableHttpTransport(url=url, headers=headers)
                oauth = auth.oauth
            case _:
                msg = f"Unsupported server config type: {type(config)}"
                raise ValueError(msg)

        # Create message handler if needed
        msg_handler: MessageHandlerT | MessageHandler | None
        if not self._message_handler:
            from llmling_agent.mcp_server.message_handler import MCPMessageHandler

            msg_handler = MCPMessageHandler(self)
        else:
            msg_handler = self._message_handler

        return fastmcp.Client(
            transport,
            log_handler=self._log_handler,
            roots=self._accessible_roots,
            timeout=config.timeout,
            elicitation_handler=self._elicitation_handler_impl,
            sampling_handler=self._sampling_handler_impl,
            message_handler=msg_handler,
            auth="oauth" if (force_oauth or oauth) else None,
        )

    async def _refresh_tools(self) -> None:
        """Refresh the list of available tools from the server."""
        if not self._client or not self._connected:
            return

        try:
            tools = await self._client.list_tools()
            self._available_tools = tools
            logger.debug("Refreshed %d tools from MCP server", len(tools))
        except Exception as e:  # noqa: BLE001
            logger.warning("Failed to refresh tools: %s", e)
            self._available_tools = []

    def get_tools(self) -> list[dict[str, Any]]:
        """Get tools in OpenAI function format."""
        return [
            {"type": "function", "function": mcp_tool_to_fn_schema(tool)}
            for tool in self._available_tools
        ]

    async def list_prompts(self) -> list[MCPPrompt]:
        """Get available prompts from the server."""
        if not self._client or not self._connected:
            msg = "Not connected to MCP server"
            raise RuntimeError(msg)

        try:
            return await self._client.list_prompts()
        except Exception as e:  # noqa: BLE001
            logger.debug("Failed to list prompts: %s", e)
            return []

    async def list_resources(self) -> list[MCPResource]:
        """Get available resources from the server."""
        if not self._client or not self._connected:
            msg = "Not connected to MCP server"
            raise RuntimeError(msg)

        try:
            return await self._client.list_resources()
        except Exception as e:
            msg = f"Failed to list resources: {e}"
            raise RuntimeError(msg) from e

    async def get_prompt(
        self, name: str, arguments: dict[str, str] | None = None
    ) -> mcp.types.GetPromptResult:
        """Get a specific prompt's content."""
        if not self._client or not self._connected:
            msg = "Not connected to MCP server"
            raise RuntimeError(msg)

        try:
            return await self._client.get_prompt_mcp(name, arguments)
        except Exception as e:
            msg = f"Failed to get prompt {name!r}: {e}"
            raise RuntimeError(msg) from e

    def convert_tool(self, tool: MCPTool) -> Tool:
        """Create a properly typed callable from MCP tool schema."""
        from llmling_agent import Tool

        schema = mcp_tool_to_fn_schema(tool)
        fn_schema = FunctionSchema.from_dict(schema)
        sig = fn_schema.to_python_signature()

        async def tool_callable(ctx: RunContext, **kwargs: Any) -> str | Any | ToolReturn:
            """Dynamically generated MCP tool wrapper."""
            # Filter out None values for optional params
            schema_props = tool.inputSchema.get("properties", {})
            required_props = set(tool.inputSchema.get("required", []))

            filtered_kwargs = {
                k: v
                for k, v in kwargs.items()
                if k in required_props or (k in schema_props and v is not None)
            }
            tc_id = ctx.tool_call_id if ctx else None

            return await self.call_tool(
                tool.name,
                filtered_kwargs,
                tool_call_id=tc_id,
                run_context=ctx,
            )

        # Set proper signature and annotations with RunContext support
        tool_callable.__signature__ = _create_tool_signature_with_context(sig)  # type: ignore
        annotations = _create_tool_annotations_with_context(fn_schema.get_annotations())
        # Update return annotation to support multiple types
        annotations["return"] = str | Any | ToolReturn
        tool_callable.__annotations__ = annotations

        tool_callable.__name__ = tool.name
        tool_callable.__doc__ = tool.description or "No description provided."
        meta = {"mcp_tool": tool.name}
        return Tool.from_callable(tool_callable, source="mcp", metadata=meta)

    def _create_final_progress_handler(
        self, tool_name: str, tool_call_id: str, tool_input: dict[str, Any]
    ) -> ProgressHandler:
        """Create a FastMCP-compatible progress handler with baked-in context."""

        async def fastmcp_progress_handler(
            progress: float, total: float | None, message: str | None
        ):
            await self._progress_handler(
                progress, total, message, tool_name, tool_call_id, tool_input
            )

        return fastmcp_progress_handler

    async def call_tool(
        self,
        name: str,
        arguments: dict | None = None,
        tool_call_id: str | None = None,
        run_context: RunContext | None = None,
    ) -> ToolReturn | str | Any:
        """Call an MCP tool with full PydanticAI return type support."""
        if not self._client or not self._connected:
            msg = "Not connected to MCP server"
            raise RuntimeError(msg)

        # Create progress handler if we have handler
        progress_handler = None
        if self._progress_handler:
            if run_context and run_context.tool_call_id and run_context.tool_name:
                # Extract tool args from message history
                tool_input = extract_tool_call_args(
                    run_context.messages, run_context.tool_call_id
                )
                progress_handler = self._create_final_progress_handler(
                    run_context.tool_name, run_context.tool_call_id, tool_input
                )
            else:
                # Fallback to using passed arguments (direct tool call)
                progress_handler = self._create_final_progress_handler(
                    name, tool_call_id or "", arguments or {}
                )
        try:
            result = await self._client.call_tool(
                name, arguments, progress_handler=progress_handler
            )
            content = await self._convert_mcp_content(result.content)
            # Decision logic for return type
            match (result.data is not None, bool(content)):
                case (True, True):  # Both structured data and rich content -> ToolReturn
                    return ToolReturn(return_value=result.data, content=content)
                case (True, False):  # Only structured data -> return directly
                    return result.data
                case (False, True):  # Only content -> ToolReturn with content
                    msg = "Tool executed successfully"
                    return ToolReturn(return_value=msg, content=content)
                case (False, False):  # Fallback to text extraction
                    return extract_text_content(result.content)
                case _:  # Handle unexpected cases
                    msg = f"Unexpected MCP content: {result.content}"
                    raise ValueError(msg)  # noqa: TRY301
        except Exception as e:
            msg = f"MCP tool call failed: {e}"
            raise RuntimeError(msg) from e

    async def _convert_mcp_content(self, mcp_content: list[Any]) -> list[str | Any]:
        """Convert MCP content blocks to PydanticAI content types."""
        from mcp.types import (
            AudioContent,
            BlobResourceContents,
            EmbeddedResource,
            ImageContent,
            ResourceLink,
            TextContent,
            TextResourceContents,
        )

        contents: list[Any] = []

        for block in mcp_content:
            match block:
                case TextContent(text=text):
                    contents.append(text)
                case TextResourceContents(text=text):
                    contents.append(text)
                case ImageContent(data=data, mimeType=mime_type):
                    decoded_data = base64.b64decode(data)
                    img = BinaryImage(data=decoded_data, media_type=mime_type)
                    contents.append(img)
                case AudioContent(data=data, mimeType=mime_type):
                    decoded_data = base64.b64decode(data)
                    content = BinaryContent(data=decoded_data, media_type=mime_type)
                    contents.append(content)
                case BlobResourceContents(blob=blob):
                    decoded_data = base64.b64decode(blob)
                    mime = "application/octet-stream"
                    content = BinaryContent(data=decoded_data, media_type=mime)
                    contents.append(content)
                case ResourceLink(uri=uri):
                    try:
                        assert self._client
                        res = await self._client.read_resource_mcp(uri)
                        if len(res.contents) == 1:
                            nested = await self._convert_mcp_content([res.contents[0]])
                            contents.extend(nested)
                        else:
                            nested = await self._convert_mcp_content(res.contents)
                            contents.extend(nested)
                    except Exception:  # noqa: BLE001
                        # Fallback to DocumentUrl if reading fails
                        contents.append(DocumentUrl(url=str(uri)))
                case EmbeddedResource(resource=TextResourceContents(text=text)):
                    contents.append(text)
                case EmbeddedResource(resource=BlobResourceContents() as blob_resource):
                    contents.append(f"[Binary data: {blob_resource.mimeType}]")
                case _:
                    contents.append(str(block))  # Convert anything else to string

        return contents

__aenter__ async

__aenter__() -> Self

Enter context manager.

Source code in src/llmling_agent/mcp_server/client.py

async def __aenter__(self) -> Self:
    """Enter context manager."""
    try:
        # First attempt with configured auth
        await self._client.__aenter__()

    except Exception as first_error:
        # OAuth fallback for HTTP/SSE if not already using OAuth
        if _should_try_oauth_fallback(self.config):
            try:
                with contextlib.suppress(Exception):
                    await self._client.__aexit__(None, None, None)
                self._client = self._get_client(self.config, force_oauth=True)
                await self._client.__aenter__()
                logger.info("Connected with OAuth fallback")
            except Exception:  # noqa: BLE001
                raise first_error from None
        else:
            raise

    self._connected = True
    await self._refresh_tools()
    return self

__aexit__ async

__aexit__(
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
)

Exit context manager and cleanup.

Source code in src/llmling_agent/mcp_server/client.py

async def __aexit__(
    self,
    exc_type: type[BaseException] | None,
    exc_val: BaseException | None,
    exc_tb: TracebackType | None,
):
    """Exit context manager and cleanup."""
    try:
        await self._client.__aexit__(None, None, None)
    except Exception as e:  # noqa: BLE001
        logger.warning("Error during FastMCP client cleanup: %s", e)
    finally:
        self._connected = False
        self._available_tools = []

call_tool async

call_tool(
    name: str,
    arguments: dict | None = None,
    tool_call_id: str | None = None,
    run_context: RunContext | None = None,
) -> ToolReturn | str | Any

Call an MCP tool with full PydanticAI return type support.

Source code in src/llmling_agent/mcp_server/client.py

async def call_tool(
    self,
    name: str,
    arguments: dict | None = None,
    tool_call_id: str | None = None,
    run_context: RunContext | None = None,
) -> ToolReturn | str | Any:
    """Call an MCP tool with full PydanticAI return type support."""
    if not self._client or not self._connected:
        msg = "Not connected to MCP server"
        raise RuntimeError(msg)

    # Create progress handler if we have handler
    progress_handler = None
    if self._progress_handler:
        if run_context and run_context.tool_call_id and run_context.tool_name:
            # Extract tool args from message history
            tool_input = extract_tool_call_args(
                run_context.messages, run_context.tool_call_id
            )
            progress_handler = self._create_final_progress_handler(
                run_context.tool_name, run_context.tool_call_id, tool_input
            )
        else:
            # Fallback to using passed arguments (direct tool call)
            progress_handler = self._create_final_progress_handler(
                name, tool_call_id or "", arguments or {}
            )
    try:
        result = await self._client.call_tool(
            name, arguments, progress_handler=progress_handler
        )
        content = await self._convert_mcp_content(result.content)
        # Decision logic for return type
        match (result.data is not None, bool(content)):
            case (True, True):  # Both structured data and rich content -> ToolReturn
                return ToolReturn(return_value=result.data, content=content)
            case (True, False):  # Only structured data -> return directly
                return result.data
            case (False, True):  # Only content -> ToolReturn with content
                msg = "Tool executed successfully"
                return ToolReturn(return_value=msg, content=content)
            case (False, False):  # Fallback to text extraction
                return extract_text_content(result.content)
            case _:  # Handle unexpected cases
                msg = f"Unexpected MCP content: {result.content}"
                raise ValueError(msg)  # noqa: TRY301
    except Exception as e:
        msg = f"MCP tool call failed: {e}"
        raise RuntimeError(msg) from e

convert_tool

convert_tool(tool: Tool) -> Tool

Create a properly typed callable from MCP tool schema.

Source code in src/llmling_agent/mcp_server/client.py

def convert_tool(self, tool: MCPTool) -> Tool:
    """Create a properly typed callable from MCP tool schema."""
    from llmling_agent import Tool

    schema = mcp_tool_to_fn_schema(tool)
    fn_schema = FunctionSchema.from_dict(schema)
    sig = fn_schema.to_python_signature()

    async def tool_callable(ctx: RunContext, **kwargs: Any) -> str | Any | ToolReturn:
        """Dynamically generated MCP tool wrapper."""
        # Filter out None values for optional params
        schema_props = tool.inputSchema.get("properties", {})
        required_props = set(tool.inputSchema.get("required", []))

        filtered_kwargs = {
            k: v
            for k, v in kwargs.items()
            if k in required_props or (k in schema_props and v is not None)
        }
        tc_id = ctx.tool_call_id if ctx else None

        return await self.call_tool(
            tool.name,
            filtered_kwargs,
            tool_call_id=tc_id,
            run_context=ctx,
        )

    # Set proper signature and annotations with RunContext support
    tool_callable.__signature__ = _create_tool_signature_with_context(sig)  # type: ignore
    annotations = _create_tool_annotations_with_context(fn_schema.get_annotations())
    # Update return annotation to support multiple types
    annotations["return"] = str | Any | ToolReturn
    tool_callable.__annotations__ = annotations

    tool_callable.__name__ = tool.name
    tool_callable.__doc__ = tool.description or "No description provided."
    meta = {"mcp_tool": tool.name}
    return Tool.from_callable(tool_callable, source="mcp", metadata=meta)

get_prompt async

get_prompt(name: str, arguments: dict[str, str] | None = None) -> GetPromptResult

Get a specific prompt's content.

Source code in src/llmling_agent/mcp_server/client.py

async def get_prompt(
    self, name: str, arguments: dict[str, str] | None = None
) -> mcp.types.GetPromptResult:
    """Get a specific prompt's content."""
    if not self._client or not self._connected:
        msg = "Not connected to MCP server"
        raise RuntimeError(msg)

    try:
        return await self._client.get_prompt_mcp(name, arguments)
    except Exception as e:
        msg = f"Failed to get prompt {name!r}: {e}"
        raise RuntimeError(msg) from e

get_tools

get_tools() -> list[dict[str, Any]]

Get tools in OpenAI function format.

Source code in src/llmling_agent/mcp_server/client.py

def get_tools(self) -> list[dict[str, Any]]:
    """Get tools in OpenAI function format."""
    return [
        {"type": "function", "function": mcp_tool_to_fn_schema(tool)}
        for tool in self._available_tools
    ]

list_prompts async

list_prompts() -> list[Prompt]

Get available prompts from the server.

Source code in src/llmling_agent/mcp_server/client.py

async def list_prompts(self) -> list[MCPPrompt]:
    """Get available prompts from the server."""
    if not self._client or not self._connected:
        msg = "Not connected to MCP server"
        raise RuntimeError(msg)

    try:
        return await self._client.list_prompts()
    except Exception as e:  # noqa: BLE001
        logger.debug("Failed to list prompts: %s", e)
        return []

list_resources async

list_resources() -> list[Resource]

Get available resources from the server.

Source code in src/llmling_agent/mcp_server/client.py

async def list_resources(self) -> list[MCPResource]:
    """Get available resources from the server."""
    if not self._client or not self._connected:
        msg = "Not connected to MCP server"
        raise RuntimeError(msg)

    try:
        return await self._client.list_resources()
    except Exception as e:
        msg = f"Failed to list resources: {e}"
        raise RuntimeError(msg) from e