haystack/docs-website/reference_versioned_docs/version-2.29/integrations-api/mcp.md at f04ba18c50942524bdf6eff5b31b697a3e6866ed · deepset-ai/haystack

title	MCP
id	integrations-mcp
description	MCP integration for Haystack
slug	/integrations-mcp

haystack_integrations.tools.mcp.mcp_tool

AsyncExecutor

Thread-safe event loop executor for running async code from sync contexts.

get_instance

get_instance() -> AsyncExecutor

Get or create the global singleton executor instance.

init

__init__() -> None

Initialize a dedicated event loop

run

run(coro: Coroutine[Any, Any, Any], timeout: float | None = None) -> Any

Run a coroutine in the event loop.

Parameters:

coro (Coroutine[Any, Any, Any]) – Coroutine to execute
timeout (float | None) – Optional timeout in seconds

Returns:

Any – Result of the coroutine

Raises:

TimeoutError – If execution exceeds timeout

get_loop

get_loop() -> asyncio.AbstractEventLoop

Get the event loop.

Returns:

AbstractEventLoop – The event loop

run_background

run_background(
    coro_factory: Callable[[asyncio.Event], Coroutine[Any, Any, Any]],
    timeout: float | None = None,
) -> tuple[concurrent.futures.Future[Any], asyncio.Event]

Schedule coro_factory to run in the executor's event loop without blocking the caller thread.

The factory receives an :class:asyncio.Event that can be used to cooperatively shut the coroutine down. The method returns both the concurrent future (to observe completion or failure) and the created stop_event so that callers can signal termination.

Parameters:

coro_factory (Callable\[[Event], Coroutine[Any, Any, Any]]) – A callable receiving the stop_event and returning the coroutine to execute.
timeout (float | None) – Optional timeout while waiting for the stop_event to be created.

Returns:

tuple[Future[Any], Event] – Tuple (future, stop_event).

shutdown

shutdown(timeout: float = 2) -> None

Shut down the background event loop and thread.

Parameters:

timeout (float) – Timeout in seconds for shutting down the event loop

MCPError

Bases: Exception

Base class for MCP-related errors.

init

__init__(message: str) -> None

Initialize the MCPError.

Parameters:

message (str) – Descriptive error message

MCPConnectionError

Bases: MCPError

Error connecting to MCP server.

init

__init__(
    message: str,
    server_info: MCPServerInfo | None = None,
    operation: str | None = None,
) -> None

Initialize the MCPConnectionError.

Parameters:

message (str) – Descriptive error message
server_info (MCPServerInfo | None) – Server connection information that was used
operation (str | None) – Name of the operation that was being attempted

MCPToolNotFoundError

Bases: MCPError

Error when a tool is not found on the server.

init

__init__(
    message: str, tool_name: str, available_tools: list[str] | None = None
) -> None

Initialize the MCPToolNotFoundError.

Parameters:

message (str) – Descriptive error message
tool_name (str) – Name of the tool that was requested but not found
available_tools (list[str] | None) – List of available tool names, if known

MCPInvocationError

Bases: ToolInvocationError

Error during tool invocation.

init

__init__(
    message: str, tool_name: str, tool_args: dict[str, Any] | None = None
) -> None

Initialize the MCPInvocationError.

Parameters:

message (str) – Descriptive error message
tool_name (str) – Name of the tool that was being invoked
tool_args (dict[str, Any] | None) – Arguments that were passed to the tool

MCPClient

Bases: ABC

Abstract base class for MCP clients.

This class defines the common interface and shared functionality for all MCP clients, regardless of the transport mechanism used.

connect

connect() -> list[types.Tool]

Connect to an MCP server.

Returns:

list[Tool] – List of available tools on the server

Raises:

MCPConnectionError – If connection to the server fails

call_tool

call_tool(tool_name: str, tool_args: dict[str, Any]) -> str

Call a tool on the connected MCP server.

Parameters:

tool_name (str) – Name of the tool to call
tool_args (dict[str, Any]) – Arguments to pass to the tool

Returns:

str – JSON string representation of the tool invocation result

Raises:

MCPConnectionError – If not connected to an MCP server
MCPInvocationError – If the tool invocation fails

aclose

aclose() -> None

Close the connection and clean up resources.

This method ensures all resources are properly released, even if errors occur.

StdioClient

Bases: MCPClient

MCP client that connects to servers using stdio transport.

init

__init__(
    command: str,
    args: list[str] | None = None,
    env: dict[str, str | Secret] | None = None,
    max_retries: int = 3,
    base_delay: float = 1.0,
    max_delay: float = 30.0,
) -> None

Initialize a stdio MCP client.

Parameters:

command (str) – Command to run (e.g., "python", "node")
args (list[str] | None) – Arguments to pass to the command
env (dict[str, str | Secret] | None) – Environment variables for the command
max_retries (int) – Maximum number of reconnection attempts
base_delay (float) – Base delay for exponential backoff in seconds

connect

connect() -> list[types.Tool]

Connect to an MCP server using stdio transport.

Returns:

list[Tool] – List of available tools on the server

Raises:

MCPConnectionError – If connection to the server fails

SSEClient

Bases: MCPClient

MCP client that connects to servers using SSE transport.

init

__init__(
    server_info: SSEServerInfo,
    max_retries: int = 3,
    base_delay: float = 1.0,
    max_delay: float = 30.0,
) -> None

Initialize an SSE MCP client using server configuration.

Parameters:

server_info (SSEServerInfo) – Configuration object containing URL, token, timeout, etc.
max_retries (int) – Maximum number of reconnection attempts
base_delay (float) – Base delay for exponential backoff in seconds

connect

connect() -> list[types.Tool]

Connect to an MCP server using SSE transport.

Note: If both custom headers and token are provided, custom headers take precedence.

Returns:

list[Tool] – List of available tools on the server

Raises:

MCPConnectionError – If connection to the server fails

StreamableHttpClient

Bases: MCPClient

MCP client that connects to servers using streamable HTTP transport.

init

__init__(
    server_info: StreamableHttpServerInfo,
    max_retries: int = 3,
    base_delay: float = 1.0,
    max_delay: float = 30.0,
) -> None

Initialize a streamable HTTP MCP client using server configuration.

Parameters:

server_info (StreamableHttpServerInfo) – Configuration object containing URL, token, timeout, etc.
max_retries (int) – Maximum number of reconnection attempts
base_delay (float) – Base delay for exponential backoff in seconds

connect

connect() -> list[types.Tool]

Connect to an MCP server using streamable HTTP transport.

Note: If both custom headers and token are provided, custom headers take precedence.

Returns:

list[Tool] – List of available tools on the server

Raises:

MCPConnectionError – If connection to the server fails

MCPServerInfo

Bases: ABC

Abstract base class for MCP server connection parameters.

This class defines the common interface for all MCP server connection types.

create_client

create_client() -> MCPClient

Create an appropriate MCP client for this server info.

Returns:

MCPClient – An instance of MCPClient configured with this server info

to_dict

to_dict() -> dict[str, Any]

Serialize this server info to a dictionary.

Returns:

dict[str, Any] – Dictionary representation of this server info

from_dict

from_dict(data: dict[str, Any]) -> MCPServerInfo

Deserialize server info from a dictionary.

Parameters:

data (dict[str, Any]) – Dictionary containing serialized server info

Returns:

MCPServerInfo – Instance of the appropriate server info class

SSEServerInfo

Bases: MCPServerInfo

Data class that encapsulates SSE MCP server connection parameters.

For authentication tokens containing sensitive data, you can use Secret objects for secure handling and serialization:

server_info = SSEServerInfo(
    url="https://my-mcp-server.com",
    token=Secret.from_env_var("API_KEY"),
)

For custom headers (e.g., non-standard authentication):

# Single custom header with Secret
server_info = SSEServerInfo(
    url="https://my-mcp-server.com",
    headers={"X-API-Key": Secret.from_env_var("API_KEY")},
)

# Multiple headers (mix of Secret and plain strings)
server_info = SSEServerInfo(
    url="https://my-mcp-server.com",
    headers={
        "X-API-Key": Secret.from_env_var("API_KEY"),
        "X-Client-ID": "my-client-id",
    },
)

Parameters:

url (str | None) – Full URL of the MCP server (including /sse endpoint)
base_url (str | None) – Base URL of the MCP server (deprecated, use url instead)
token (str | Secret | None) – Authentication token for the server (optional, generates "Authorization: Bearer <token>" header)
headers (dict[str, str | Secret] | None) – Custom HTTP headers (optional, takes precedence over token parameter if provided)
timeout (int) – Connection timeout in seconds

create_client

create_client() -> MCPClient

Create an SSE MCP client.

Returns:

MCPClient – Configured MCPClient instance

StreamableHttpServerInfo

Bases: MCPServerInfo

Data class that encapsulates streamable HTTP MCP server connection parameters.

For authentication tokens containing sensitive data, you can use Secret objects for secure handling and serialization:

server_info = StreamableHttpServerInfo(
    url="https://my-mcp-server.com",
    token=Secret.from_env_var("API_KEY"),
)

For custom headers (e.g., non-standard authentication):

# Single custom header with Secret
server_info = StreamableHttpServerInfo(
    url="https://my-mcp-server.com",
    headers={"X-API-Key": Secret.from_env_var("API_KEY")},
)

# Multiple headers (mix of Secret and plain strings)
server_info = StreamableHttpServerInfo(
    url="https://my-mcp-server.com",
    headers={
        "X-API-Key": Secret.from_env_var("API_KEY"),
        "X-Client-ID": "my-client-id",
    },
)

Parameters:

url (str) – Full URL of the MCP server (streamable HTTP endpoint)
token (str | Secret | None) – Authentication token for the server (optional, generates "Authorization: Bearer <token>" header)
headers (dict[str, str | Secret] | None) – Custom HTTP headers (optional, takes precedence over token parameter if provided)
timeout (int) – Connection timeout in seconds

create_client

create_client() -> MCPClient

Create a streamable HTTP MCP client.

Returns:

MCPClient – Configured StreamableHttpClient instance

StdioServerInfo

Bases: MCPServerInfo

Data class that encapsulates stdio MCP server connection parameters.

Parameters:

command (str) – Command to run (e.g., "python", "node")
args (list[str] | None) – Arguments to pass to the command
env (dict[str, str | Secret] | None) – Environment variables for the command

For environment variables containing sensitive data, you can use Secret objects for secure handling and serialization:

server_info = StdioServerInfo(
    command="uv",
    args=["run", "my-mcp-server"],
    env={
        "WORKSPACE_PATH": "/path/to/workspace",  # Plain string
        "API_KEY": Secret.from_env_var("API_KEY"),  # Secret object
    }
)

Secret objects will be properly serialized and deserialized without exposing the secret value, while plain strings will be preserved as-is. Use Secret objects for sensitive data that needs to be handled securely.

create_client

create_client() -> MCPClient

Create a stdio MCP client.

Returns:

MCPClient – Configured StdioMCPClient instance

MCPTool

Bases: Tool

A Tool that represents a single tool from an MCP server.

This implementation uses the official MCP SDK for protocol handling while maintaining compatibility with the Haystack tool ecosystem.

Response handling:

Text and image content are supported and returned as JSON strings
The JSON contains the structured response from the MCP server
Use json.loads() to parse the response into a dictionary

State-mapping support:

MCPTool supports state-mapping parameters (outputs_to_string, inputs_from_state, outputs_to_state)
These enable integration with Agent state for automatic parameter injection and output handling
See the __init__ method documentation for details on each parameter

Example using Streamable HTTP:

import json
from haystack_integrations.tools.mcp import MCPTool, StreamableHttpServerInfo

# Create tool instance
tool = MCPTool(
    name="multiply",
    server_info=StreamableHttpServerInfo(url="http://localhost:8000/mcp")
)

# Use the tool and parse result
result_json = tool.invoke(a=5, b=3)
result = json.loads(result_json)

Example using SSE (deprecated):

import json
from haystack.tools import MCPTool, SSEServerInfo

# Create tool instance
tool = MCPTool(
    name="add",
    server_info=SSEServerInfo(url="http://localhost:8000/sse")
)

# Use the tool and parse result
result_json = tool.invoke(a=5, b=3)
result = json.loads(result_json)

Example using stdio:

import json
from haystack.tools import MCPTool, StdioServerInfo

# Create tool instance
tool = MCPTool(
    name="get_current_time",
    server_info=StdioServerInfo(command="python", args=["path/to/server.py"])
)

# Use the tool and parse result
result_json = tool.invoke(timezone="America/New_York")
result = json.loads(result_json)

init

__init__(
    name: str,
    server_info: MCPServerInfo,
    description: str | None = None,
    connection_timeout: int = 30,
    invocation_timeout: int = 30,
    eager_connect: bool = False,
    outputs_to_string: dict[str, Any] | None = None,
    inputs_from_state: dict[str, str] | None = None,
    outputs_to_state: dict[str, dict[str, Any]] | None = None,
) -> None

Initialize the MCP tool.

Parameters:

name (str) – Name of the tool to use
server_info (MCPServerInfo) – Server connection information
description (str | None) – Custom description (if None, server description will be used)
connection_timeout (int) – Timeout in seconds for server connection
invocation_timeout (int) – Default timeout in seconds for tool invocations
eager_connect (bool) – If True, connect to server during initialization. If False (default), defer connection until warm_up or first tool use, whichever comes first.
outputs_to_string (dict[str, Any] | None) – Optional dictionary defining how tool outputs should be converted into a string. If the source is provided only the specified output key is sent to the handler. If the source is omitted the whole tool result is sent to the handler. Example: {"source": "docs", "handler": my_custom_function}
inputs_from_state (dict[str, str] | None) – Optional dictionary mapping state keys to tool parameter names. Example: {"repository": "repo"} maps state's "repository" to tool's "repo" parameter.
outputs_to_state (dict[str, dict[str, Any]] | None) – Optional dictionary defining how tool outputs map to keys within state as well as optional handlers. If the source is provided only the specified output key is sent to the handler. Example with source: {"documents": {"source": "docs", "handler": custom_handler}} Example without source: {"documents": {"handler": custom_handler}}

Raises:

MCPConnectionError – If connection to the server fails
MCPToolNotFoundError – If no tools are available or the requested tool is not found
TimeoutError – If connection times out

ainvoke

ainvoke(**kwargs: Any) -> str | dict[str, Any]

Asynchronous tool invocation.

Parameters:

kwargs (Any) – Arguments to pass to the tool

Returns:

str | dict[str, Any] – JSON string or dictionary representation of the tool invocation result. Returns a dictionary when outputs_to_state is configured to enable state updates.

Raises:

MCPInvocationError – If the tool invocation fails
TimeoutError – If the operation times out

warm_up

warm_up() -> None

Connect and fetch the tool schema if eager_connect is turned off.

to_dict

to_dict() -> dict[str, Any]

Serializes the MCPTool to a dictionary.

The serialization preserves all information needed to recreate the tool, including server connection parameters, timeout settings, and state-mapping parameters. Note that the active connection is not maintained.

Returns:

dict[str, Any] – Dictionary with serialized data in the format: {"type": fully_qualified_class_name, "data": {parameters}}

from_dict

from_dict(data: dict[str, Any]) -> Tool

Deserializes the MCPTool from a dictionary.

This method reconstructs an MCPTool instance from a serialized dictionary, including recreating the server_info object and state-mapping parameters. A new connection will be established to the MCP server during initialization.

Parameters:

data (dict[str, Any]) – Dictionary containing serialized tool data

Returns:

Tool – A fully initialized MCPTool instance

Raises:

Exception – if connection fails

close

close() -> None

Close the tool synchronously.

haystack_integrations.tools.mcp.mcp_toolset

MCPToolset

Bases: Toolset

A Toolset that connects to an MCP (Model Context Protocol) server and provides access to its tools.

MCPToolset dynamically discovers and loads all tools from any MCP-compliant server, supporting both network-based streaming connections (Streamable HTTP, SSE) and local process-based stdio connections. This dual connectivity allows for integrating with both remote and local MCP servers.

Example using MCPToolset in a Haystack Pipeline:

# Prerequisites:
# 1. pip install uvx mcp-server-time  # Install required MCP server and tools
# 2. export OPENAI_API_KEY="your-api-key"  # Set up your OpenAI API key

import os
from haystack import Pipeline
from haystack.components.converters import OutputAdapter
from haystack.components.generators.chat import OpenAIChatGenerator
from haystack.components.tools import ToolInvoker
from haystack.dataclasses import ChatMessage
from haystack_integrations.tools.mcp import MCPToolset, StdioServerInfo

# Create server info for the time service (can also use SSEServerInfo for remote servers)
server_info = StdioServerInfo(command="uvx", args=["mcp-server-time", "--local-timezone=Europe/Berlin"])

# Create the toolset - this will automatically discover all available tools
# You can optionally specify which tools to include
mcp_toolset = MCPToolset(
    server_info=server_info,
    tool_names=["get_current_time"]  # Only include the get_current_time tool
)

# Create a pipeline with the toolset
pipeline = Pipeline()
pipeline.add_component("llm", OpenAIChatGenerator(model="gpt-4o-mini", tools=mcp_toolset))
pipeline.add_component("tool_invoker", ToolInvoker(tools=mcp_toolset))
pipeline.add_component(
    "adapter",
    OutputAdapter(
        template="{{ initial_msg + initial_tool_messages + tool_messages }}",
        output_type=list[ChatMessage],
        unsafe=True,
    ),
)
pipeline.add_component("response_llm", OpenAIChatGenerator(model="gpt-4o-mini"))
pipeline.connect("llm.replies", "tool_invoker.messages")
pipeline.connect("llm.replies", "adapter.initial_tool_messages")
pipeline.connect("tool_invoker.tool_messages", "adapter.tool_messages")
pipeline.connect("adapter.output", "response_llm.messages")

# Run the pipeline with a user question
user_input = "What is the time in New York? Be brief."
user_input_msg = ChatMessage.from_user(text=user_input)

result = pipeline.run({"llm": {"messages": [user_input_msg]}, "adapter": {"initial_msg": [user_input_msg]}})
print(result["response_llm"]["replies"][0].text)

You can also use the toolset via Streamable HTTP to talk to remote servers:

from haystack_integrations.tools.mcp import MCPToolset, StreamableHttpServerInfo

# Create the toolset with streamable HTTP connection
toolset = MCPToolset(
    server_info=StreamableHttpServerInfo(url="http://localhost:8000/mcp"),
    tool_names=["multiply"]  # Optional: only include specific tools
)
# Use the toolset as shown in the pipeline example above

Example with state configuration for Agent integration:

from haystack_integrations.tools.mcp import MCPToolset, StdioServerInfo

# Create the toolset with per-tool state configuration
# This enables tools to read from and write to the Agent's State
toolset = MCPToolset(
    server_info=StdioServerInfo(command="uvx", args=["mcp-server-git"]),
    tool_names=["git_status", "git_diff", "git_log"],

    # Maps the state key "repository" to the tool parameter "repo_path" for each tool
    inputs_from_state={
        "git_status": {"repository": "repo_path"},
        "git_diff": {"repository": "repo_path"},
        "git_log": {"repository": "repo_path"},
    },
    # Map tool outputs to state keys for each tool
    outputs_to_state={
        "git_status": {"status_result": {"source": "status"}},  # Extract "status" from output
        "git_diff": {"diff_result": {}},  # use full output with default handling
    },
)

Example using SSE (deprecated):

from haystack_integrations.tools.mcp import MCPToolset, SSEServerInfo
from haystack.components.tools import ToolInvoker

# Create the toolset with an SSE connection
sse_toolset = MCPToolset(
    server_info=SSEServerInfo(url="http://some-remote-server.com:8000/sse"),
    tool_names=["add", "subtract"]  # Only include specific tools
)

# Use the toolset as shown in the pipeline example above

init

__init__(
    server_info: MCPServerInfo,
    tool_names: list[str] | None = None,
    connection_timeout: float = 30.0,
    invocation_timeout: float = 30.0,
    eager_connect: bool = False,
    inputs_from_state: dict[str, dict[str, str]] | None = None,
    outputs_to_state: dict[str, dict[str, dict[str, Any]]] | None = None,
    outputs_to_string: dict[str, dict[str, Any]] | None = None,
) -> None

Initialize the MCP toolset.

Parameters:

server_info (MCPServerInfo) – Connection information for the MCP server
tool_names (list[str] | None) – Optional list of tool names to include. If provided, only tools with matching names will be added to the toolset.
connection_timeout (float) – Timeout in seconds for server connection
invocation_timeout (float) – Default timeout in seconds for tool invocations
eager_connect (bool) – If True, connect to server and load tools during initialization. If False (default), defer connection to warm_up.
inputs_from_state (dict[str, dict[str, str]] | None) – Optional dictionary mapping tool names to their inputs_from_state config. Each config maps state keys to tool parameter names. Tool names should match available tools from the server; a warning is logged for unknown tools. Note: With Haystack >= 2.22.0, parameter names are validated; ValueError is raised for invalid parameters. With earlier versions, invalid parameters fail at runtime. Example: {"git_status": {"repository": "repo_path"}}
outputs_to_state (dict[str, dict[str, dict[str, Any]]] | None) – Optional dictionary mapping tool names to their outputs_to_state config. Each config defines how tool outputs map to state keys with optional handlers. Tool names should match available tools from the server; a warning is logged for unknown tools. Example: {"git_status": {"status_result": {"source": "status"}}}
outputs_to_string (dict[str, dict[str, Any]] | None) – Optional dictionary mapping tool names to their outputs_to_string config. Each config defines how tool outputs are converted to strings. Tool names should match available tools from the server; a warning is logged for unknown tools. Example: {"git_diff": {"source": "diff", "handler": format_diff}}

Raises:

MCPToolNotFoundError – If any of the specified tool names are not found on the server
ValueError – If parameter names in inputs_from_state are invalid (Haystack >= 2.22.0 only)

warm_up

warm_up() -> None

Connect and load tools when eager_connect is turned off.

This method is automatically called by ToolInvoker.warm_up() and Pipeline.warm_up(). You can also call it directly before using the toolset to ensure all tool schemas are available without performing a real invocation.

to_dict

to_dict() -> dict[str, Any]

Serialize the MCPToolset to a dictionary.

Returns:

dict[str, Any] – A dictionary representation of the MCPToolset

from_dict

from_dict(data: dict[str, Any]) -> MCPToolset

Deserialize an MCPToolset from a dictionary.

Parameters:

data (dict[str, Any]) – Dictionary representation of the MCPToolset

Returns:

MCPToolset – A new MCPToolset instance

close

close() -> None

Close the underlying MCP client safely.

FilesExpand file tree

mcp.md

Latest commit

History

mcp.md

File metadata and controls

haystack_integrations.tools.mcp.mcp_tool

AsyncExecutor

get_instance

init

run

get_loop

run_background

shutdown

MCPError

init

MCPConnectionError

init

MCPToolNotFoundError

init

MCPInvocationError

init

MCPClient

connect

call_tool

aclose

StdioClient

init

connect

SSEClient

init

connect

StreamableHttpClient

init

connect

MCPServerInfo

create_client

to_dict

from_dict

SSEServerInfo

create_client

StreamableHttpServerInfo

create_client

StdioServerInfo

create_client

MCPTool

init

ainvoke

warm_up

to_dict

from_dict

close

haystack_integrations.tools.mcp.mcp_toolset

MCPToolset

init

warm_up

to_dict

from_dict

close