LFX - Langflow Executor

The Langflow Executor (LFX) is a command-line tool that serves and runs flows statelessly from flow JSON files with minimal dependencies.

Running a flow with LFX is similar to running flows with the --backend-only environment variable enabled, but even more lightweight because the Langflow package and all of its dependencies don't need to be installed.

LFX uses a no-op database interface called NoopSession for all operations that require persistent state. There is no langflow.db database file when using LFX. You can run flows with the API, but any stateful operations that depend on the Langflow database, like saving flows, storing messages, or managing users will not persist data. Operations that depend on langflow.db will not work as they do in the full Langflow application.

Memory operations are dispatched at call time, not at import time. If the langflow package is installed in the same Python environment as lfx and a real database service is registered, memory operations are routed to the full langflow.memory implementation instead. This applies when lfx is used as a Python library inside a running Langflow server, not when running lfx run or lfx serve from the command line.

Commands

Runtime commands — documented in this README:

Flow DevOps SDK commands — documented in the Flow DevOps Toolkit:

Prerequisites

• Install Python.

• Install uv.

• Create or download a flow JSON file. For example, download the Simple Agent flow from the repository:

  bashCopy

  curl -o simple-agent-flow.json "https://raw.githubusercontent.com/langflow-ai/langflow/main/src/backend/base/langflow/initial_setup/starter_projects/Simple%20Agent.json"
  

• Create an OpenAI API key.

• Create a Langflow API key. For LFX, you can generate a secure token locally (see Serve the simple agent starter flow with lfx serve), or create one through the Langflow server UI or CLI.

Install LFX

LFX can be installed in multiple ways. If you have installed Langflow OSS version >=1.6, lfx is already included.

Clone repository

1. Clone the Langflow repository:

   bashCopy

   git clone https://github.com/langflow-ai/langflow
   

2. Change directory to langflow/src/lfx:

   bashCopy

   cd langflow/src/lfx
   

   From this directory, you can run lfx commands using uv run lfx as shown in lfx serve or lfx run.

Install from PyPI

1. Create and activate a virtual environment:

   bashCopy

   uv venv lfx-venv
   source lfx-venv/bin/activate
   

2. Install the LFX package from PyPI:

   bashCopy

   uv pip install lfx
   

   To install the latest nightly (pre-release) version of LFX:

   bashCopy

   uv pip install --pre lfx
   

   To run lfx commands, continue to lfx serve or lfx run.

Run without installing

Run LFX without installing it locally using uvx.

1. Create a Langflow API key (see Serve), and set LANGFLOW_API_KEY in the same terminal session as lfx:

   bashCopy

   export LANGFLOW_API_KEY="sk..."
   

2. Run lfx serve using uvx:

   bashCopy

   uvx lfx serve simple-agent-flow.json
   

   This command downloads and runs LFX in a temporary environment without permanent installation. From the same environment, you can also run flows directly with lfx run.

Serve the simple agent starter flow with lfx serve

lfx serve starts a FastAPI server that hosts one or more flows. You can load flows at startup from files or a directory, or start with an empty registry and upload flows via the API. Once running, flows are available at POST /flows/{flow_id}/run.

lfx serve accepts a .json flow file or a .py Python script (same as lfx run), as well as inline JSON via --flow-json or piped input via --stdin.

lfx serve accepts a .json flow file or a .py Python script (same as lfx run), as well as inline JSON via --flow-json or piped input via --stdin.

The API key is required for security because lfx serve can create a publicly accessible FastAPI server.

This example uses the Agent component's built-in OpenAI model, which requires an OpenAI API key. If you want to use a different provider, edit the model provider, model name, and credentials accordingly.

1. Generate a Langflow API key.

   For LFX, you can generate a secure token locally to use as your LANGFLOW_API_KEY:

   bashCopy

   uv run python -c "import secrets; print(secrets.token_urlsafe(32))"
   

   This is different from creating a Langflow API key through the Langflow server UI or CLI, which stores the key in the Langflow database. For LFX, you only need a secure token string to authenticate requests to your LFX server.

2. Set up your environment variables using one of the following options.

   Option: .env file

   Create a .env file and populate it with your flow's variables. The LANGFLOW_API_KEY is required. This example assumes the flow requires an OpenAI API key.

   bashCopy

   LANGFLOW_API_KEY="sk..."
   OPENAI_API_KEY="sk-..."
   

   Option: Export variables

   Export your variables in the same terminal session where you'll start the server. You must declare your variables before the server starts for the server to pick them up.

   bashCopy

   export LANGFLOW_API_KEY="sk..."
   export OPENAI_API_KEY="sk-..."
   

3. Install dependencies.

   If you already have Langflow installed, or if you're running from source at src/lfx, LFX is included with Langflow and all dependencies are already available. You don't need to install additional dependencies.

   If you install the standalone lfx package from PyPI or run LFX with uvx, you need to manually install the dependencies required by the components in your flow.

   To find which dependencies your flow requires:

   1. Run your flow with lfx run:

      bashCopy

      uv run lfx run simple-agent-flow.json "test input"
      

      LFX reports any missing dependencies in the subsequent error message.

   2. Install the missing dependencies that LFX reports.

   For example, to run the simple agent template flow, install these dependencies in your environment before running the simple agent flow:

   bashCopy

   uv pip install "langchain~=0.3.23" "langchain-core<1.0.0" "langchain-community" "langchain-openai" "langchain-text-splitters" beautifulsoup4 lxml requests
   

4. Start the server with your variable values using one of the following options.

   Option: .env file

   This example assumes your flow file and .env file are in the current directory:

   bashCopy

   uv run lfx serve simple-agent-flow.json --env-file .env
   

   If your .env file is in a different location, provide the full or relative path:

   bashCopy

   uv run lfx serve simple-agent-flow.json --env-file /path/to/.env
   

   Option: Export variables

   If you exported your variables, the command to start the server automatically picks up the values when it starts:

   bashCopy

   uv run lfx serve simple-agent-flow.json
   

   To export new values, stop the server, export the variables, and then start the server again.

5. The startup process displays a flow_id value in the output. Copy the flow_id to use in the test API call in the next step. In this example, the flow_id is c1dab29d-3364-58ef-8fef-99311d32ee42:

   Copy

    LFX Server
    Flow loaded: simple-agent-flow.json (c1dab29d-3364-58ef-8fef-99311d32ee42)
    Server:      http://127.0.0.1:8000
    Run flows at: POST /flows/{flow_id}/run
    API key:     x-api-key header or ?x-api-key= query parameter
   

6. To send a test request to the server, open a new terminal and export your flow_id and Langflow API key values as variables:

   bashCopy

   export LANGFLOW_API_KEY="sk..."
   export FLOW_ID="c1dab29d-3364-58ef-8fef-99311d32ee42"
   

7. Test the server with an API call to the /flows/flow_id/run endpoint:

   bashCopy

   curl -X POST http://localhost:8000/flows/$FLOW_ID/run \
     -H "Content-Type: application/json" \
     -H "x-api-key: $LANGFLOW_API_KEY" \
     -d '{"input_value": "Hello, world!"}'
   

   Successful response example:

   jsonCopy

   {
     "result": "Hello world! 👋\n\nHow can I help you today? If you have any questions or need assistance, just let me know!",
     "success": true,
     "logs": "\n\n\u001b[1m> Entering new None chain...\u001b[0m\n\u001b[32;1m\u001b[1;3mHello world! 👋\n\nHow can I help you today?...\u001b[0m\n\n\u001b[1m> Finished chain.\u001b[0m\n",
     "type": "message",
     "component": "Chat Output"
   }
   

Your flow is now running as a lightweight API endpoint, with only the flow's required dependencies and no visual builder installed. Users who call your endpoint don't need to install Langflow or configure their own LLM provider keys.

To make your server publicly accessible, use a tunneling service like ngrok or deploy to a public cloud provider.

HTTP endpoints

The LFX server exposes the following endpoints. All /flows/{flow_id} routes require the x-api-key header or ?x-api-key= query parameter.

Request body schema

POST /flows/{flow_id}/run

{
  "input_value": "Your message here",
  "session_id": "optional-conversation-id"
}

session_id is optional. When set, Agent and Memory components use it to maintain conversation history across multiple calls. If omitted, a new session ID is generated for each request.

POST /flows/{flow_id}/stream

{
  "input_value": "Your message here",
  "input_type": "chat",
  "output_type": "chat",
  "output_component": null,
  "session_id": "optional-conversation-id",
  "tweaks": {"ComponentName": {"param": "value"}}
}

Response schema

The LFX server's response schema is different from the Langflow API /run endpoint's schema. Requests to the LFX server's /flows/{flow_id}/run endpoint return the following fields:

{
  "result": "string",      // Output result from the flow execution
  "success": true,         // Whether execution was successful
  "logs": "string",        // Captured logs from execution
  "type": "message",       // Type of result
  "component": "string"    // The component that generated the result (e.g. "Chat Output")
}

The /stream endpoint returns the same fields as server-sent events, with one event per component output.

Serve multiple flows at startup

You can pass a directory, multiple file paths, or a mix of both to load several flows at startup. Each flow is registered under its own ID.

# Serve every .json file in a directory (top-level only, not recursive)
uv run lfx serve flows/

# Serve specific files
uv run lfx serve flow-a.json flow-b.json

# Mix directory and individual files
uv run lfx serve flows/ extra-flow.json

Python .py script flows are also supported when using a single worker:

uv run lfx serve my_flow.py

Start with no flows

lfx serve starts with an empty registry when no flow path is provided. Flows can then be uploaded via the API:

uv run lfx serve --env-file .env

Upload flows dynamically

While the server is running, upload flows with POST /flows/upload/.

The endpoint accepts the full Langflow export JSON directly, the same format you get from the Langflow UI's Export button:

# Upload a flow JSON file
curl -X POST http://localhost:8000/flows/upload/ \
  -H "x-api-key: $LANGFLOW_API_KEY" \
  -H "Content-Type: application/json" \
  -d @my-flow.json

# Upload and replace an existing flow with the same ID
curl -X POST "http://localhost:8000/flows/upload/?replace=true" \
  -H "x-api-key: $LANGFLOW_API_KEY" \
  -H "Content-Type: application/json" \
  -d @my-flow.json

The server returns a 409 Conflict error if you upload a flow whose ID is already registered and replace=true is not set.

Multi-worker deployment

Use --workers N to start multiple uvicorn worker processes, and --flow-dir to point all workers at a shared directory so they serve the same flows.

# 4 workers, flows shared via a local temp directory
uv run lfx serve flows/ --workers 4 --flow-dir /tmp/lfx-flows

# Cross-pod sharing (Kubernetes PVC or similar network volume)
uv run lfx serve flows/ --workers 4 --flow-dir /mnt/shared-flows

How it works:

• Each startup flow file is persisted to --flow-dir as {flow_id}.json when the server starts.
• Uploads via POST /flows/upload/ are also written to --flow-dir, making them visible to every worker on the next request.
• Deletes propagate across workers: each worker detects a missing file on the next request to that flow and returns 404.
• Without --flow-dir, each worker maintains an isolated in-memory registry. Uploads reach only the worker that received the request.

Note: --workers > 1 with --flow-dir does not support .py script flows — Python graphs cannot be serialized to the filesystem store.

A warning is printed when --workers > 1 is used without --flow-dir, because in that case each worker has its own isolated in-memory registry.

Isolate credentials with --no-env-fallback (experimental)

By default, lfx serve resolves component credentials from the process environment (os.environ). In multi-tenant deployments, this means every request shares the same credentials.

Use --no-env-fallback to disable process-environment fallback. With this flag set, credentials must be supplied per-request in the global_vars field of the request body:

uv run lfx serve my-flow.json --no-env-fallback --env-file .env

Pass per-request credentials in the global_vars map under LANGFLOW_REQUEST_VARIABLES:

curl -X POST http://localhost:8000/flows/$FLOW_ID/run \
  -H "Content-Type: application/json" \
  -H "x-api-key: $LANGFLOW_API_KEY" \
  -d '{
    "input_value": "Hello world",
    "global_vars": {
      "LANGFLOW_REQUEST_VARIABLES": {
        "OPENAI_API_KEY": "sk-per-request-key"
      }
    }
  }'

Credentials supplied in LANGFLOW_REQUEST_VARIABLES are scoped to the current request using Python contextvars. Langflow's built-in components do not write them to os.environ, so they do not bleed into other concurrent requests on the same worker. Custom components that explicitly write to os.environ are outside this guarantee.

Check or upgrade flow compatibility at startup

Use --upgrade-flow to check compatibility between a flow and the current LFX version before serving it. See LFX and Langflow version compatibility for details on the version model.

# Fail at startup if any component is incompatible
uv run lfx serve my-flow.json --upgrade-flow=check

# Apply safe upgrades in memory, then start serving
uv run lfx serve my-flow.json --upgrade-flow=safe

LFX serve options

Run the simple agent flow with lfx run

The lfx run command runs a flow from a JSON file without serving it, and the output is sent to stdout. Input to lfx run can be a path to the JSON file, inline JSON passed with --input-value, or read from stdin. No Langflow API key is required.

This example uses the Agent component's built-in OpenAI model, which requires an OpenAI API key. If you want to use a different provider, edit the model provider, model name, and credentials accordingly.

1. Export your variables in the same terminal session where you'll run the flow:

   bashCopy

   export OPENAI_API_KEY="sk-..."
   

2. Install dependencies.

   If you already have Langflow installed, or if you're running from source at src/lfx, LFX is included with Langflow and all dependencies are already available. You don't need to install additional dependencies.

   If you install the standalone lfx package from PyPI or run LFX with uvx, you need to manually install the dependencies required by the components in your flow.

   To find which dependencies your flow requires:

   1. Run your flow with lfx run:

      bashCopy

      uv run lfx run simple-agent-flow.json "test input"
      

      LFX reports any missing dependencies in the subsequent error message.

   2. Install the missing dependencies that LFX reports.

   For example, to run the simple agent template flow, install these dependencies in your environment before running the simple agent flow:

   bashCopy

   uv pip install "langchain~=0.3.23" "langchain-core<1.0.0" "langchain-community" "langchain-openai" "langchain-text-splitters" beautifulsoup4 lxml requests
   

3. Run the flow from a flow JSON file:

   bashCopy

   uv run lfx run simple-agent-flow.json "Hello world"
   

   This flow expects a Message input, which is a simple text string.

   You can also use the --input-value flag instead of a positional argument:

   bashCopy

   uv run lfx run simple-agent-flow.json --input-value "Hello world"
   

   The --input-value flag is required when using --stdin or --flow-json options, since those options use the positional argument for the flow definition instead of the input value.

In addition to running flows from JSON files, lfx run supports other input methods, described below.

Run flows from stdin

The --stdin option lets you run flows from dynamic sources (APIs, databases) or after modifying a flow before execution. The command reads the flow's JSON definition from stdin, validates the JSON structure, and runs the flow. The --input-value flag is required when using --stdin.

Read a flow JSON from stdin:

cat simple-agent-flow.json | uv run lfx run --stdin \
  --input-value "Hello world" \
  --format json | jq '.result'

Fetch a flow JSON from a remote API and run it:

curl https://api.example.com/flows/my-agent-flow | uv run lfx run --stdin \
  --input-value "Hello world"

Modify a flow created in the visual builder before execution (e.g. change the OpenAI model to gpt-4o):

cat simple-agent-flow.json | jq '(.data.nodes[] | select(.data.node.template.model_name.value) | .data.node.template.model_name.value) = "gpt-4o"' | \
  uv run lfx run --stdin \
  --input-value "Hello world" \
  --format json | jq '.result'

Run flows with inline JSON

Instead of piping from stdin or reading from a JSON file, you can pass the flow JSON directly as a string argument. The --input-value flag is required when using --flow-json.

uv run lfx run --flow-json '{"data": {"nodes": [...], "edges": [...]}}' \
  --input-value "Hello world"

LFX run options

In addition to running flows from JSON files, you can use lfx run with Python scripts that define flows programmatically. This approach allows you to create flows directly in Python code without the visual builder.

For a complete example of creating an agent flow programmatically using LFX components, see the Complete Agent Example on PyPI or the Complete Agent Example below.

Complete agent example

Create a file called simple_agent.py:

"""A simple agent flow example for Langflow.

Usage:
    uv run lfx run simple_agent.py "How are you?"
"""

import os
from pathlib import Path

from lfx import components as cp
from lfx.graph import Graph
from lfx.log.logger import LogConfig


async def get_graph() -> Graph:
    """Create and return the graph with async component initialization."""
    log_config = LogConfig(
        log_level="INFO",
        log_file=Path("langflow.log"),
    )

    chat_input = cp.ChatInput()
    agent = cp.AgentComponent()
    url_component = cp.URLComponent()
    tools = await url_component.to_toolkit()

    agent.set(
        model_name="gpt-4.1-mini",
        agent_llm="OpenAI",
        api_key=os.getenv("OPENAI_API_KEY"),
        input_value=chat_input.message_response,
        tools=tools,
    )
    chat_output = cp.ChatOutput().set(input_value=agent.message_response)

    return Graph(chat_input, chat_output, log_config=log_config)

Install dependencies and set your OpenAI API key, then run:

uv run lfx run simple_agent.py "How are you?" --verbose

lfx-mcp

lfx-mcp is a separate binary installed with lfx. It starts an MCP server that gives any MCP-compatible client programmatic control over a Langflow instance for building flows, managing components, and running executions.

For more information, see LFX_MCP.md.

Development

# Install development dependencies
make dev

# Run tests
make test

# Format code
make format

Pluggable services

LFX supports a pluggable service architecture that lets you customize and extend its behavior. You can replace built-in services (storage, telemetry, tracing, etc.) with your own implementations or use Langflow's full-featured services.

For more information, see PLUGGABLE_SERVICES.md.

Flattened component access

LFX supports simplified component imports for a better developer experience when building flows in Python. You get simpler imports, easier discovery via cp.ComponentName, and full backward compatibility with the traditional import method.

Before (old import style):

from lfx.components.agents.agent import AgentComponent
from lfx.components.data.url import URLComponent
from lfx.components.input_output import ChatInput, ChatOutput

Now (flattened style):

from lfx import components as cp

chat_input = cp.ChatInput()
agent = cp.AgentComponent()
url_component = cp.URLComponent()
chat_output = cp.ChatOutput()

Component category allowlist and blocklist

You can restrict which component categories are available when loading flows by using an allowlist or a blocklist.

Environment variables

Both settings are optional. When unset or empty, all categories from the component index are loaded.

Category names are case-insensitive.

Component categories

Category names in the allowlist and blocklist match the component index (e.g. top-level folders under lfx.components). The virtual keyword core in the allowlist or blocklist expands to the following core categories (aligned with the frontend sidebar):

• input_output, data_source, models_and_agents, llm_operations, files_and_knowledge, processing, flow_controls, utilities, prototypes, tools, agents, data, logic, helpers, models, vectorstores, inputs, outputs, prompts, chains, documentloaders, link_extractors, output_parsers, retrievers, textsplitters, toolkits

Provider-specific and other categories (e.g. openai, anthropic, google, langchain_utilities) are also valid; the full set depends on your LFX version and index.

How to use in LFX

1. Set one or both environment variables before running lfx serve or lfx run. The filter is applied when the component index is loaded.

Allowlist only — restrict to specific categories:

export LANGFLOW_COMPONENT_CATEGORY_ALLOWLIST="openai,anthropic,google,processing,input_output"
uv run lfx serve my_flow.json

Blocklist only — load all categories except the ones you exclude:

export LANGFLOW_COMPONENT_CATEGORY_BLOCKLIST="prototypes,langchain_utilities"
uv run lfx run my_flow.json "Hello"

Virtual core keyword — use core in the allowlist or blocklist to refer to all core categories at once (e.g. allow only core categories, or exclude all core from a broader set):

export LANGFLOW_COMPONENT_CATEGORY_ALLOWLIST="core"
uv run lfx serve my_flow.json

License

MIT License. See LICENSE for details.