Skip to main content

Model Context Protocol (MCP) server

Langflow integrates with the Model Context Protocol (MCP) as both an MCP server and an MCP client. This page describes how to use Langflow as an MCP server. For information about using Langflow as an MCP client, see the MCP connection component.

As an MCP server, Langflow exposes your flows as tools that MCP clients can use use to take actions.

Prerequisites​

  • A Langflow project with at least one flow created.

  • Any LTS version of Node.js installed on your computer to use MCP Inspector to test and debug flows.

Select and configure flows to expose as MCP tools​

Langflow runs a separate MCP server for every project. The MCP server for each project exposes that project's flows as tools.

All of the flows in a project are exposed by default. To expose only specific flows and optionally rename them for agentic use, follow these steps:

  1. From the Langflow dashboard, select the project that contains the flows you want to serve as tools, and then click the MCP Server tab. Alternatively, you can quickly access the MCP Server tab from within any flow by selecting Publish > MCP Server.

    The MCP Server tab displays a code template that you can use to connect MCP clients to the the project's MCP server.

    The Flows/Actions section lists the flows that are currently being served as tools.

    MCP server projects page

  2. Click Edit Actions.

  3. In the MCP Server Actions window, select the flows that you want exposed as tools.

    MCP server actions

  4. Optional: Edit the Flow Name and Flow Description.

    • Flow Name: Enter a name thats makes it clear what the flow does.

    • Flow Description: Enter a description that accurately describes the specific action(s) the flow performs.

    important

    MCP clients use the Flow Name and Flow Description to determine which action to use. For more information about naming and describing your flows, see Name and describe your flows for agentic use.

  5. Close the MCP Server Actions window to save your changes.

Connect clients to Langflow's MCP server​

The following procedure describes how to connect Cursor to your Langflow project's MCP server to consume your flows as tools. However, you can connect any MCP-compatible client following similar steps.

  1. Install Cursor.

  2. In Cursor, go to Cursor Settings > MCP, and then click Add New Global MCP Server. This opens Cursor's global MCP configuration file, mcp.json.

  3. In the Langflow dashboard, select the project that contains the flows you want to serve, and then click the MCP Server tab.

  4. Copy the code template from the MCP Server tab, and then paste it into mcp.json in Cursor. For example:


    _11
    {
    _11
    "mcpServers": {
    _11
    "PROJECT_NAME": {
    _11
    "command": "uvx",
    _11
    "args": [
    _11
    "mcp-proxy",
    _11
    "http://LANGFLOW_SERVER_ADDRESS/api/v1/mcp/project/PROJECT_ID/sse"
    _11
    ]
    _11
    }
    _11
    }
    _11
    }

    The MCP Server tab automatically includes the correct PROJECT_NAME, LANGFLOW_SERVER_ADDRESS, and PROJECT_ID values. The default Langflow server address is http://127.0.0.1:7860 (http://127.0.0.1:7868 if using Langflow for Desktop).

    important

    If your Langflow server requires authentication (LANGFLOW_AUTO_LOGIN is set to false), you must include your Langflow API key in the configuration. For more information, see MCP server authentication and environment variables.

  5. Save and close the mcp.json file in Cursor. The newly added MCP server will appear in the MCP Servers section.

Cursor is now connected to your project's MCP server and your flows are registered as tools. Cursor determines when to use tools based on your queries, and requests permissions when necessary.

For more information, see the Cursor's MCP documentation.

MCP server authentication and environment variables​

If your Langflow server requires authentication (LANGFLOW_AUTO_LOGIN is set to false), then you must supply a Langflow API key in your MCP client configuration. When this is the case, the code template in your project's MCP Server tab automatically includes the --header and x-api-key arguments:


_14
{
_14
"mcpServers": {
_14
"PROJECT_NAME": {
_14
"command": "uvx",
_14
"args": [
_14
"mcp-proxy",
_14
"--headers",
_14
"x-api-key",
_14
"YOUR_API_KEY",
_14
"http://LANGFLOW_SERVER_ADDRESS/api/v1/mcp/project/PROJECT_ID/sse"
_14
]
_14
}
_14
}
_14
}

Click Generate API key to automatically insert a new Langflow API key into the code template. Alternatively, you can replace YOUR_API_KEY with an existing Langflow API key.

MCP server tab showing Generate API key button

To include environment variables with your MCP server command, include them like this:


_14
{
_14
"mcpServers": {
_14
"PROJECT_NAME": {
_14
"command": "uvx",
_14
"args": [
_14
"mcp-proxy",
_14
"http://LANGFLOW_SERVER_ADDRESS/api/v1/mcp/project/PROJECT_ID/sse"
_14
],
_14
"env": {
_14
"KEY": "VALUE"
_14
}
_14
}
_14
}
_14
}

Replace KEY and VALUE with the environment variable name and value you want to include.

Name and describe your flows for agentic use​

MCP clients like Cursor "see" your Langflow project as a single MCP server, with all of your enabled flows listed as tools. This can confuse agents. For example, an agent won't know that flow adbbf8c7-0a34-493b-90ea-5e8b42f78b66 is a Document Q&A flow for a specific text file.

To prevent this behavior, make sure to name and describe your flows clearly. It's helpful to think of the names and descriptions as function names and code comments, making sure to use clear statements describing the problems your flows solve.

For example, let's say you have a Document Q&A flow that loads a sample resume for an LLM to chat with, and that you've given it the following name and description:

  • Flow Name: document_qa_for_resume

  • Flow Description: A flow for analyzing Emily's resume.

If you ask Cursor a question specifically about the resume, such as What job experience does Emily have?, the agent asks to call the MCP tool document_qa_for_resume. That's because your name and description provided the agent with a clear purpose for the tool.

When you run the tool, the agent requests permissions when necessary, and then provides a response. For example:


_10
{
_10
"input_value": "What job experience does Emily have?"
_10
}
_10
Result:
_10
What job experience does Emily have?
_10
Emily J. Wilson has the following job experience:

If you ask about a different resume, such as What job experience does Alex have?, you've provided enough information in the description for the agent to make the correct decision:


_10
I notice you're asking about Alex's job experience.
_10
Based on the available tools, I can see there is a Document QA for Resume flow that's designed for analyzing resumes.
_10
However, the description mentions it's for "Emily's resume" not Alex's. I don't have access to Alex's resume or job experience information.

Use MCP Inspector to test and debug flows​

MCP Inspector is a common tool for testing and debugging MCP servers. You can use MCP Inspector to monitor your flows and get insights into how they are being consumed by the MCP server:

  1. Install MCP Inspector:


    _10
    npx @modelcontextprotocol/inspector

    For more information about configuring MCP Inspector, including specifying a proxy port, see the MCP Inspector GitHub project.

  2. Open a web browser and navigate to the MCP Inspector UI. The default address is http://127.0.0.1:6274.

  3. In the MCP Inspector UI, enter the connection details for your Langflow project's MCP server:

    • Transport Type: Select SSE.
    • URL: Enter the Langflow MCP server's sse endpoint. For example: http://127.0.0.1:7860/api/v1/mcp/project/d359cbd4-6fa2-4002-9d53-fa05c645319c/sse

    If you've configured authentication for your MCP server, fill out the following additional fields:

    • Transport Type: Select STDIO.
    • Command: uvx
    • Arguments: Enter the following list of arguments, separated by spaces. Replace the values for YOUR_API_KEY, LANGFLOW_SERVER_ADDRESS, and PROJECT_ID with the values from your Langflow MCP server. For example:

    _10
    mcp-proxy --headers x-api-key YOUR_API_KEY http://LANGFLOW_SERVER_ADDRESS/api/v1/mcp/project/PROJECT_ID/sse

  4. Click Connect.

    If the connection was successful, you should see your project's flows in the Tools tab. From this tab, you can monitor how your flows are being registered as tools by MCP, as well as test the tools with custom input values.

  5. To quit MCP Inspector, press Control+C in the same terminal window where you started it.

Deploy your MCP server externally​

By default, Langflow isn't exposed to the public internet. However, you can forward Langflow server traffic with a forwarding platform like ngrok or zrok.

The following procedure uses ngrok, but you can use any similar reverse proxy or forwarding platform. This procedure also assumes that you're using the default Langflow listening address http://127.0.0.1:7860 (http://127.0.0.1:7868 if using Langflow for Desktop).

  1. Sign up for an ngrok account.

  2. Install ngrok.

  3. Copy your ngrok authtoken and use it to authenticate your local ngrok server:


    _10
    ngrok config add-authtoken NGROK_TOKEN

    Replace NGROK_TOKEN with your ngrok authtoken.

  4. Use ngrok to expose your Langflow server to the public internet:


    _10
    ngrok http http://localhost:7860

    The ngrok session starts in your terminal and deploys an ephemeral domain with no authentication. To add authentication or deploy a static domain, see the ngrok documentation.

    The Forwarding row displays the forwarding address for your Langflow server:


    _10
    Forwarding https://94b1-76-64-171-14.ngrok-free.app -> http://localhost:7860

    The forwarding address is acting as a reverse proxy for your Langflow server.

  5. From the Langflow dashboard, select the project that contains the flows you want to serve as tools, and then click the MCP Server tab.

    Note that the code template now contains your ngrok forwarding address instead of the localhost address:


    _11
    {
    _11
    "mcpServers": {
    _11
    "PROJECT_NAME": {
    _11
    "command": "uvx",
    _11
    "args": [
    _11
    "mcp-proxy",
    _11
    "https://94b1-73-64-171-14.ngrok-free.app/api/v1/mcp/project/fdbc12af-0dd4-43dc-b9ce-c324d1ce5cd1/sse"
    _11
    ]
    _11
    }
    _11
    }
    _11
    }

  6. Complete the steps in Connect clients to Langflow's MCP server using the ngrok forwarding address.

Your MCP client is now connected to your project's MCP server over the public internet.

If using Cursor, your conversations are the same as they are on your local host:


_10
{
_10
"input_value": "What job experience does Emily have?"
_10
}
_10
Result:
_10
What job experience does Emily have?
_10
Emily J. Wilson has the following job experience:

You can use the ngrok console output to monitor requests for your project's endpoint:


_10
16:35:48.566 EDT GET /api/v1/mcp/project/fdbc12af-0dd4-43dc-b9ce-c324d1ce5cd1 200 OK

Troubleshooting MCP server​

If Claude for Desktop is not using your server's tools correctly, you may need to explicitly define the path to your local uvx or npx executable file in the claude_desktop_config.json configuration file.

  1. To find your UVX path, run which uvx. To find your NPX path, run which npx.

  2. Copy the path, and then replace PATH_TO_UVX or PATH_TO_NPX in your claude_desktop_config.json file.


_11
{
_11
"mcpServers": {
_11
"PROJECT_NAME": {
_11
"command": "PATH_TO_UVX",
_11
"args": [
_11
"mcp-proxy",
_11
"http://LANGFLOW_SERVER_ADDRESS/api/v1/mcp/project/PROJECT_ID/sse"
_11
]
_11
}
_11
}
_11
}

Search