Langflow

Langflow is an open-source, visual workflow builder designed to make it easy to design, prototype, and run LLM-powered applications without losing architectural rigor. Instead of wiring prompts, models, tools, retrievers, and agents together purely in code, Langflow lets you compose these building blocks as a graph—making complex AI systems easier to reason about, debug, and evolve.

Langflow dramatically shortens the path from idea to working AI application while still producing production-ready artifacts: flows can be versioned, exported as JSON, executed via APIs, and integrated into real systems. For teams experimenting with RAG, agentic workflows, tool calling, or MCP servers, Langflow provides a shared visual language that aligns developers, architects, and product stakeholders

How does Langflow relate to what you’ve built so far? In previous modules you wrote Python scripts to create agents, configure tools, and invoke MCP servers. Langflow provides the same capabilities through a visual drag-and-drop interface. The flow you’ll build here is functionally equivalent to your Python agents — it connects to the same vLLM models and MCP servers — but makes the architecture visible and easier to iterate on.

Note: This is a fairly advanced exercise as it involves the building of the agent from scratch.

Prerequites

You have the FantaCo Backend and MCP Servers up and running.

export CUSTOMER_MCP_SERVER_URL=https://$(oc get routes -l app=mcp-customer -o jsonpath="{range .items[*]}{.status.ingress[0].host}{end}")/mcp
echo $CUSTOMER_MCP_SERVER_URL

export FINANCE_MCP_SERVER_URL=https://$(oc get routes -l app=mcp-finance -o jsonpath="{range .items[*]}{.status.ingress[0].host}{end}")/mcp
echo $FINANCE_MCP_SERVER_URL

https://mcp-customer-route-agentic-user1.apps.ocp.zr7nb.sandbox2823.opentlc.com/mcp
https://mcp-finance-route-agentic-user1.apps.ocp.zr7nb.sandbox2823.opentlc.com/mcp

At this end of this module, you should have something that looks similiar to the following:

Setup

cd $HOME/fantaco-redhat-one-2026/langflow-setup/

oc apply -f langflow-openshift.yaml

persistentvolumeclaim/langflow-data created
deployment.apps/langflow created
service/langflow created
route.route.openshift.io/langflow created

oc get pods -w

NAME                                           READY   STATUS              RESTARTS       AGE
fantaco-customer-main-7fd4ddb666-pfm82         1/1     Running             0              29m
fantaco-finance-main-75ffddb44b-v487g          1/1     Running             0              29m
langflow-57c45bb775-pvghq                      0/1     ContainerCreating   0              26s
langfuse-clickhouse-shard0-0                   1/1     Running             0              7m32s
langfuse-postgresql-0                          1/1     Running             0              7m32s
langfuse-redis-primary-0                       1/1     Running             0              7m32s
langfuse-s3-5fb6c8f845-mxlhh                   1/1     Running             0              7m32s
langfuse-web-6474bb4b98-shjtm                  1/1     Running             3 (7m6s ago)   7m32s
langfuse-worker-55f7bd48c4-jmwj5               1/1     Running             0              7m32s
langfuse-zookeeper-0                           1/1     Running             0              7m32s
langgraph-fastapi-569d6d554-97vtv              1/1     Running             0              29m
llamastack-distribution-vllm-67b6cb888-zj72c   1/1     Running             0              31m
mcp-customer-6bd8bcfc7b-25w7r                  1/1     Running             0              29m
mcp-finance-8cc684b8d-st274                    1/1     Running             0              29m
my-workbench-0                                 2/2     Running             0              13m
postgresql-customer-ff78dffdf-d478q            1/1     Running             0              29m
postgresql-finance-689d97894f-qrch2            1/1     Running             0              29m
simple-agent-chat-ui-6d7794dc6b-4kngc          1/1     Running             0              29m

It can take several minutes to bring up the langflow pod

Wait for 1/1 and Running

export LANGFLOW_URL="https://$(oc get route langflow -o jsonpath='{.spec.host}')"
echo $LANGFLOW_URL

Hello World

Create a new flow - Click Create first flow

Click Blank Flow

Drag a Text Input component onto the canvas

Drag a Chat Output component onto the canvas

Connect them together (drag from the output node to the input node)

Provide a message Hello Aloha Bonjour

Click the Playground button

Using this icon to get back to the list of all projects and flows

You can use the ellipses … on the Project or Flow to make changes such as its name

vLLM chatbot

+ New Flow

+ Blank Flow

Langflow does NOT have an out-of-the-box (OOTB) Component that works with vLLM via MaaS where you need to override:

API URL
API Key
Model Name

We have a custom component. Click +New Custom Component

Click Code

Delete the current code

Copy the code from this link

vllm_model_component.py

Paste and then click Check & Save

Enter the URL, Model Name and API Key

Here are your vLLM MaaS endpoint, API key.

vLLM MaaS URL

{litellm_api_base_url}

vLLM MaaS API Key

{litellm_virtual_key}

Model Name

qwen3-14b

And change Language Model to Response (we will use Language Model later)

Add a Chat Input and Chat Output and connect the dots

Click Playground

what model are you?

Why a custom component? Langflow includes built-in components for OpenAI and other hosted providers, but vLLM served via MaaS requires a custom API URL, API key, and model name. The custom component wraps these configuration details so that vLLM appears as a native Langflow building block. Once created, it can be reused across any number of flows.

Agent

Remove Chat Input and Chat Output (for now) by single-clicking them and hitting the delete key on the keyboard. Or selecting "Delete" from the ellipsis menu.

Find "Agent" in the list of Components

Change the output of the vLLM Model Component to be Language Model

Add an Agent

And

Click on Model Provider

+ Connect other models

Awaiting model input…

Connect the vLLM Model to Agent

Add Chat Input and Chat Output to the Agent

Add MCP Component for Customer

Drag MCP Tools Component to the canvas

Click on Select a server…

Click + Add MCP Server

Select Streamable HTTP/SSE

Name: Customer

Streamable HTTP/SSE URL

export CUSTOMER_MCP_SERVER_URL=https://$(oc get routes -l app=mcp-customer -o jsonpath="{range .items[*]}{.status.ingress[0].host}{end}")/mcp
echo $CUSTOMER_MCP_SERVER_URL

Click Add Server

Toggle Tool Mode

When Response switches to Toolset, you can then connect it to the Agents Tools input

Playground and test with

who does Thomas Hardy work for?

Remember, you are dealing with a LLM and it non-determinstic behavior. In some cases, this query MAY respond with:

"Thomas Hardy was an English novelist and poet, best known for his works such as Tess of the d’Urbervilles and Far from the Madding Crowd. He was a prolific writer during the 19th and early 20th centuries. As he passed away in 1928, he does not work for anyone today. If you’re referring to a different person named Thomas Hardy, feel free to clarify!"

You can be more explicit with something like

what about the customer Thomas Hardy?

Keep going, we will work on this behavior in a bit.

Why did the agent confuse the customer with the author? Without a system prompt, the agent has no context that "Thomas Hardy" refers to a customer record. The LLM defaults to its training data, which knows Thomas Hardy as a famous English novelist. This demonstrates why system prompts are essential for domain-specific agents — you’ll fix this behavior later in the module by adding agent instructions.

Add MCP Component for Finance

+ Add MCP Server

Select Streamable HTTP/SSE

Name: Finance

Streamable HTTP/SSE URL

export FINANCE_MCP_SERVER_URL=https://$(oc get routes -l app=mcp-finance -o jsonpath="{range .items[*]}{.status.ingress[0].host}{end}")/mcp
echo $FINANCE_MCP_SERVER_URL

Click Add Server

Toggle Tool Mode

Connect MCP Servers to Agent Tools input

What are the orders for Thomas Hardy?

To help the Agent out, especially when using smaller open models, you need a better System Prompt

Add Agent Instructions

You are a helpful customer service assistant.

IMPORTANT: When ANY person's name is mentioned, ALWAYS search for them as a customer first before answering.
IMPORTANT: You MUST use tools to answer questions about customers or orders. Never guess or use general knowledge.

TOOLS AND THEIR EXACT PARAMETERS:
- search_customers(contact_name="Name") - search by customer name
- get_customer(customer_id="ID") - get customer details
- fetch_order_history(customer_id="ID") - get orders
- fetch_invoice_history(customer_id="ID") - get invoices

WORKFLOW for customer questions:
1. Call search_customers(contact_name="<customer name>")
2. Extract customer_id from results
3. Use that customer_id for other queries

Always use tools first. Never answer from general knowledge about people.

Anatomy of a good agent system prompt: This prompt does three things: (1) establishes the agent’s role ("customer service assistant"), (2) sets behavioral rules ("ALWAYS search as a customer first", "MUST use tools"), and (3) documents the available tools with their exact parameter names. Listing tool signatures in the prompt helps smaller models call tools correctly — larger models can usually figure it out from the MCP tool descriptions alone.

And now go give your agent another round of tests

What you’ve accomplished

You’ve successfully built a production-ready agentic AI application using a visual, no-code/low-code approach. Starting with a simple text input/output flow, you progressed to creating a custom vLLM component that connects to your OpenShift-hosted models, integrated multiple MCP servers as business tools (customer and finance data), and crafted effective system prompts to guide agent behavior. This visual workflow is functionally equivalent to the Python-based agents you built earlier in the workshop, but with the added benefits of easier debugging, faster iteration, and a shared visual language that bridges technical and non-technical stakeholders. Langflow demonstrates that sophisticated AI systems—complete with model inference, tool calling, and agentic reasoning—can be composed, tested, and deployed without sacrificing architectural rigor or production readiness. The flow you created can be versioned, exported as JSON, executed via API, and integrated into real enterprise systems, making it a valuable tool for both rapid prototyping and production deployment.