Tools

The langchain.mcp namespace requires langchain[mcp]>=1.4.0 and is in beta. The API may change.

MCPAdapter bridges MCP servers and LangChain agents: it discovers the tools a server advertises and adapts them into standard LangChain tools. Pass the tools from list_tools() to create_agent as you would any other LangChain tool. This page covers what is specific to that bridge: identifying MCP tools, controlling their execution, handling their outputs, and responding when a server needs input during a call. For the runnable discovery-and-agent example, see the MCP quickstart.

Use MCP tools in an agent

Discover the server’s catalog with MCPAdapter.list_tools, then give the returned tools to create_agent. From the agent’s perspective, they behave like LangChain tools: the model chooses a tool, LangChain invokes it, and the resulting ToolMessage returns to the model.

from langchain.agents import create_agent
from langchain.mcp import MCPAdapter


async def run_agent(server) -> dict:
    # Discover the server's tools, then hand them to the agent like any other
    # LangChain tools. The tools hold the client, so the agent stays usable
    # for the life of the adapter context.
    async with MCPAdapter(server) as adapter:
        tools = await adapter.list_tools()
        agent = create_agent("claude-sonnet-5", tools)
        return await agent.ainvoke(
            {
                "messages": [
                    {"role": "user", "content": "What is the forecast for Oslo?"}
                ]
            }
        )

View example trace

Open a public LangSmith run for this example.

For general guidance on defining, binding, and using LangChain tools, see Tools. For several MCP servers and their namespaced tool catalogs, see Connections.

Handle tool outputs

MCP tool results become ToolMessage objects with content the model can read, an artifact for application data, and a status indicating success or failure.

Multimodal content

The adapter converts model-visible MCP content into LangChain content blocks. For example, a tool that returns an MCP image block with a screenshot reaches the model as an image block.

from langchain.agents import create_agent
from langchain.mcp import MCPAdapter
from langchain.messages import ToolMessage


async def access_multimodal_tool_content(server) -> dict:
    async with MCPAdapter(server) as adapter:
        tools = await adapter.list_tools()
        agent = create_agent("claude-sonnet-5", tools)
        result = await agent.ainvoke(
            {"messages": [{"role": "user", "content": "Take a screenshot."}]}
        )

    # An MCP result arrives as LangChain content blocks. Image and file content
    # convert into standardized `image`/`file` blocks alongside `text`.
    for message in result["messages"]:
        if not isinstance(message, ToolMessage):
            continue
        for block in message.content_blocks:  
            if block["type"] == "text":  
                print(f"Text: {block['text']}")  
            elif block["type"] == "image":  
                preview = block.get("base64", "")[:20]  
                print(f"Image mime type: {block.get('mime_type')}")  
                print(f"Image base64: {preview}...")  

    return result

Structured content

When a tool returns structured content, the adapter attaches it to the ToolMessage as an artifact rather than folding it into the model-visible text. Run the agent, then read the artifact from the ToolMessage instances in the result:

from langchain.agents import create_agent
from langchain.mcp import MCPAdapter
from langchain.messages import ToolMessage


async def run_agent_structured(server) -> dict:
    async with MCPAdapter(server) as adapter:
        tools = await adapter.list_tools()
        agent = create_agent("claude-sonnet-5", tools)
        result = await agent.ainvoke(
            {"messages": [{"role": "user", "content": "Look up user 42."}]}
        )

    # The adapter sets `artifact` only when the tool returned structured
    # content, so a non-None artifact always carries `structured_content`.
    for message in result["messages"]:
        if isinstance(message, ToolMessage) and message.artifact is not None:
            structured = message.artifact["structured_content"]  
            print(f"Structured content: {structured}")  

    return result

The artifact is an MCPToolArtifact, whose structured_content field holds the tool result’s structuredContent. A tool that returns no structured content leaves artifact as None.

Errors

An MCP tool result carries an isError flag. When a server reports isError=True, the adapter converts it into a ToolMessage with status="error" carrying the server’s own message, so the agent can read it and correct itself:

from langchain.agents import create_agent
from langchain.mcp import MCPAdapter
from langchain.messages import ToolMessage


async def divide_by_zero(server) -> dict:
    async with MCPAdapter(server) as adapter:
        tools = await adapter.list_tools()
        agent = create_agent("claude-sonnet-5", tools)
        result = await agent.ainvoke(
            {
                "messages": [
                    {
                        "role": "user",
                        "content": "Use the divide tool to calculate 10 divided by 0.",
                    }
                ]
            }
        )

    # A server error (isError=True) reaches the model as a failed ToolMessage,
    # so the agent can read the server's own message and recover. Transport
    # failures still raise, because a model cannot act on those.
    for message in result["messages"]:
        if isinstance(message, ToolMessage) and message.status == "error":
            print(f"Tool reported: {message.text}")  

    return result

A server-reported error reaches the model as a failed tool message, but a transport or session failure raises instead.

Tool metadata

Each adapted tool may carry its MCP provenance under an mcp namespace on the LangChain tool’s metadata:

tool.metadata
# {
#     "mcp": {
#         "tool": {
#             "annotations": {
#                 "destructive_hint": True,
#                 "read_only_hint": False,
#             },
#             "_meta": {"origin": "crm"},
#         },
#         "server": {
#             "name": "crm",
#             "version": "2.1.0",
#         },
#     },
# }

Every nested field is optional: a server may provide tool annotations, _meta, server identity, any combination of those, or none. annotations contains MCP hints such as read_only_hint and destructive_hint; _meta is opaque metadata supplied by the server; and server identifies the MCP implementation that advertised the tool. Read optional metadata defensively, so a missing field returns a default rather than failing:

from langchain.tools import BaseTool


def is_destructive(tool: BaseTool) -> bool:
    """Read the MCP destructive hint off the adapter's tool metadata."""
    # Chain `.get` with defaults so a tool missing any nested field returns
    # False rather than raising.
    annotations = (
        (tool.metadata or {}).get("mcp", {}).get("tool", {}).get("annotations", {})
    )
    return annotations.get("destructive_hint", False)

Human-in-the-loop

Reading annotations lets you gate a tool based on what the server declares about it, rather than hardcoding tool names. The MCP annotation classifies the tool and LangChain’s human-in-the-loop middleware enforces the approval policy. Read the destructive hint from metadata once during tool discovery. Then give the human-in-the-loop configuration a when predicate that receives each pending tool call and returns whether the call needs approval:

from langchain.agents.middleware import HumanInTheLoopMiddleware
from langchain.agents.middleware.human_in_the_loop import InterruptOnConfig
from langchain.tools import BaseTool
from langchain.tools.tool_node import ToolCallRequest


def is_destructive(tool: BaseTool) -> bool:
    """Read the MCP destructive hint off the adapter's tool metadata."""
    annotations = (
        (tool.metadata or {}).get("mcp", {}).get("tool", {}).get("annotations", {})
    )
    return annotations.get("destructive_hint", False)


async def gate_destructive_tools(server):
    async with MCPAdapter(server) as adapter:
        tools = await adapter.list_tools()

        # Read the destructive hint from metadata once, then let a callable
        # decide per call. One config covers whatever destructive tools a
        # server exposes, without hardcoding tool names.
        destructive = {tool.name for tool in tools if is_destructive(tool)}

        def needs_approval(request: ToolCallRequest) -> bool:
            return request.tool_call["name"] in destructive

        gate = InterruptOnConfig(
            allowed_decisions=["approve", "reject"], when=needs_approval
        )
        interrupt_on: dict[str, bool | InterruptOnConfig] = {
            tool.name: gate for tool in tools
        }
        return create_agent(
            "claude-sonnet-5",
            tools,
            middleware=[HumanInTheLoopMiddleware(interrupt_on=interrupt_on)],
            checkpointer=InMemorySaver(),
        )

When the agent calls a tool that the predicate gates, the run pauses. Approve the call to run it, or reject it to skip the tool and tell the model:

from langgraph.types import Command

# Approve the pending destructive call and resume.
resumed = await agent.ainvoke(
    Command(resume={"decisions": [{"type": "approve"}]}), config
)

The predicate can also inspect the call’s arguments. This lets a tool run freely for safe inputs and pause only for risky ones, such as a delete_file call targeting a protected path. Access the arguments through request.tool_call["args"]. Combine the metadata classification and argument check to gate a tool only when its type and inputs warrant approval. For the full approval workflow, see Human-in-the-loop.

Server requests during tool execution

Most tools finish without asking the client for anything mid-call. When a server needs input, MCPAdapter surfaces elicitation as a LangGraph interrupt.

Elicitation

Elicitation lets an MCP server request input during a tool call. The adapter pauses the run with a LangGraph interrupt. Your application presents the request to the user and resumes the run with their answer. Attach a checkpointer to the agent and use the same thread_id when invoking and resuming it. This example assumes the booking tool asks for one date and pauses once. Replace the sample date with the user’s answer.

from typing import Any

from langchain.agents import create_agent
from langchain.mcp import MCPAdapter
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.types import Command


async def book_with_elicitation(server) -> dict:
    # Elicitation is handled automatically: when a server needs input mid-call,
    # the adapter surfaces the question as a LangGraph `interrupt()`, so the
    # person already reviewing the agent's work answers it and the run resumes.
    async with MCPAdapter(server) as adapter:
        tools = await adapter.list_tools()

        # Resuming a paused run needs persistence, so the interrupted run has
        # somewhere to wait.
        agent = create_agent("claude-sonnet-5", tools, checkpointer=InMemorySaver())
        config: Any = {"configurable": {"thread_id": "booking-1"}}

        paused = await agent.ainvoke(
            {"messages": [{"role": "user", "content": "Book a table for 4."}]}, config
        )
        [interrupt] = paused["__interrupt__"]
        [question] = interrupt.value["requests"]

        # Answers are keyed by the server's own request key, so nothing has to
        # be tracked across the pause. `decline` or `cancel` would refuse.
        answer = {"action": "accept", "content": {"date": "2026-09-14"}}
        return await agent.ainvoke(
            Command(resume={"responses": {question["key"]: answer}}), config
        )

Elicitation is on by default. The adapter advertises the capability and drives the interrupt loop. A prebuilt client that already carries its own elicitation handler is honored instead of overridden. Resume with Command(resume={"responses": {key: answer}}). The interrupt payload and answer types live in langchain.mcp.elicitation.

Resuming reruns the tool from the beginning. Any work performed before the server asks for input can repeat. Make that work safe to repeat without duplicating side effects.

Each answer uses one of these actions:

accept: Provide form content matching the request’s schema, or confirm completion of a URL interaction without content.
decline: Refuse to provide the requested information.
cancel: Indicate that the user canceled the interaction.

The adapter forwards the answer to the server, which decides how the tool call finishes. Only elicitation is answered this way. A server that instead asks for sampling or roots raises an error, because the modern, sessionless protocol has no live back-channel for those requests. See Sampling and roots.

Interrupt-driven elicitation answers a server that returns its request as an InputRequiredResult. A server that only pushes elicitation over a legacy handshake session cannot be answered this way.

Get started

Core components

Middleware

Frontend

Advanced usage

Agent development

Production

Use MCP tools in an agent

View example trace

Handle tool outputs

Multimodal content

Structured content

Errors

Tool metadata

Human-in-the-loop

Server requests during tool execution

Elicitation

See also

​Use MCP tools in an agent

View example trace

​Handle tool outputs

​Multimodal content

​Structured content

​Errors

​Tool metadata

​Human-in-the-loop

​Server requests during tool execution

​Elicitation

​See also

Use MCP tools in an agent

Handle tool outputs

Multimodal content

Structured content

Errors

Tool metadata

Human-in-the-loop

Server requests during tool execution

Elicitation

See also