Langflow Tutorial 2026: Build Your First Flow

This Langflow tutorial gets you from a fresh install to a working flow you can call from your own code: install Langflow, run the Simple Agent template, build a retrieval (RAG) flow over your own file, then serve it through the Langflow API. Langflow is an open-source visual builder where you drag components such as a chat input, a prompt, a model and a vector store onto a canvas, wire them together and run the result as an app.

Steps checked in October 2026 against the Langflow 1.12 docs: installation, quickstart, components, the vector RAG tutorial, triggering flows with the API and import and export. Langflow renames menu labels between releases; if a button differs, check the docs for your version.

Most beginner guides stop at the Playground. That is where Langflow is fun but not yet useful. The step that turns a canvas experiment into something a product can use is the last one, exporting the flow behind an API key and calling it from an app, so this guide treats it as part of the tutorial rather than an appendix. If you are still mapping where visual LLM builders sit next to agent frameworks and automation tools, our agentic AI workflows hub gives the wider picture first.

What you will have at the end

By the end you will have three things running on your machine: the Simple Agent template answering questions with a calculator and a URL tool, a two-part RAG flow that answers questions from a document you uploaded, and a short script that sends a question to that flow over HTTP and prints the reply. The whole sequence looks like this:

Langflow tutorial for beginners: what do you need first?

You need three accounts or tools before you open the canvas, and none of them are Langflow itself. Langflow is the wiring; the intelligence comes from a model provider, and retrieval needs an embedding model and somewhere to store vectors.

Requirements above come from the Langflow install page (checked October 2026). Pick a test document you know well. When you later check whether the RAG answers are right, you want to be able to spot a wrong one without reading the source again.

Step 1: install and start Langflow

Langflow offers three install routes, and the docs mark Langflow Desktop as the recommended one for most people. Pick based on how you plan to use it, not on which looks easiest today.

• Langflow Desktop. A standalone app for macOS and Windows that manages dependencies and upgrades for you. The docs note that a few features, such as the Shareable Playground and Voice Mode, are not available in Desktop. Good for learning on your own machine.
• Docker. Run docker run -p 7860:7860 -e LANGFLOW_AUTO_LOGIN=false -e LANGFLOW_SUPERUSER_PASSWORD=YOUR_PASSWORD langflowai/langflow:latest, replacing the password with a strong one, then open http://localhost:7860/. This is the closest to how you will run it on a server later.
• Python package. Create a virtual environment with uv venv, activate it, run uv pip install langflow, then start it with uv run langflow run. The default local address is http://127.0.0.1:7860. The docs warn that the first start can take a few minutes.

Expected result: the Langflow projects page opens in your browser (or in the Desktop window). If you chose the Python route and the page does not load, wait a little longer before assuming it failed; a slow first start is normal while dependencies load.

How Langflow components and ports fit together

Before building anything, spend two minutes on the component model, because nearly every beginner error in Langflow is a wiring error. Components are the building blocks of a flow, like classes in an application. You drag them in from the Core components or Bundles menus. Each one has inputs, outputs and parameters, and the inspection panel on the right shows the advanced settings that are hidden on the card itself.

Around each component are small circular ports. A port either accepts or emits one data type, and the colour tells you which: Message ports are indigo, Embeddings emerald, LanguageModel fuchsia, Tool cyan, JSON red. The rule from the components docs is simple: connect an output port to an input port of the same type. If two components disagree, put a processing component such as Type Convert between them.

Three controls on the component header save time later. Run component runs that component on its own. Inspect shows its output and logs. Freeze keeps the last output of that component and everything upstream, so repeat runs skip work you already paid for, which is handy once an embedding step is done.

Step 2: run your first flow from the Simple Agent template

Do not start from a blank canvas. A template shows you a correctly wired flow, which is the fastest way to learn what valid connections look like.

The Playground matters more than it looks. It shows which tool the agent picked and why, so you can see whether a wrong answer came from the model or from a tool that returned nothing. Provider details: the quickstart notes you can add only one key per provider, so that key needs access to every model you plan to use, and your provider account must have credits (checked October 2026).

Langflow RAG tutorial: chat with your own documents

Retrieval augmented generation means the model answers from your documents rather than only from what it learned in training. Langflow's Vector Store RAG template ships that pattern as two flows on one canvas, and understanding why there are two is the key to the whole tutorial.

• Load Data Flow. Read File, Split Text, an Embedding Model and a vector store. It ingests your file, cuts it into chunks, embeds each chunk and writes the chunks and vectors into the database. It only needs to run when your data changes.
• Retriever Flow. Chat Input, an Embedding Model, the same vector store, a Parser, a Prompt, a Language Model and Chat Output. It embeds each question, finds the stored chunks that sit closest to it, turns them back into text and hands them to the model as context.

Two more checks keep RAG answers honest. First, ask a question your document does not cover; a good prompt tells the model to say it does not know, and you can edit the Prompt component to say exactly that. Second, if answers cite the wrong section, adjust the chunk size and overlap in Split Text before you reach for a bigger model. The model can only use the chunks retrieval hands it.

When many people will add documents, or you want to load data from code, the RAG tutorial shows the API route: upload the file to /api/v2/files/, take the returned path, then run the Load Data Flow through /api/v1/run/FLOW_ID with that path passed as a tweak to the file component (checked October 2026).

Step 4: call your flow from an app with the Langflow API

Langflow is a visual editor, but it is also a runtime. Every flow you build can be run over HTTP, which is what makes it useful in a real product. Since Langflow 1.5, most API endpoints require a Langflow API key.

A minimal request looks like this, using the request body from the quickstart:

curl --request POST \
  --url "http://127.0.0.1:7860/api/v1/run/FLOW_ID" \
  --header "Content-Type: application/json" \
  --header "x-api-key: $LANGFLOW_API_KEY" \
  --data '{"input_type": "chat", "output_type": "chat", "input_value": "What does the manual say about servicing?"}'

Two request fields are worth knowing early. session_id can be any string; reuse it to continue a conversation thread, or leave it out and Langflow defaults it to the flow ID. tweaks are one-time overrides of a component's parameters for a single request, such as a different file path or a setting on Chat Input, without changing the saved flow. You choose which fields appear in the generated snippet by clicking API on them in the component's Parameters panel. The docs also describe a v2 Workflow API in beta, an embedded chat widget, and serving flows as an MCP server, all from the publishing docs (checked October 2026).

Your laptop is fine for building. For an app that real users hit, the docs are direct: deploy a stable Langflow server to handle API calls. That usually means the Docker image on a small server with a strong superuser password and auto-login off, which is how the official Docker command already starts it.

Export, back up and move a flow

A flow is a JSON file of nodes, edges and metadata. Export one from the Projects page (More, then Export) or from Share, then Export, and import it by dragging the JSON file onto any Langflow window. That is how you back up work, move a flow from Desktop to a server, or share it with a collaborator.

Watch the API key option. If you typed a literal key into a component and export with Save with my API keys, the key itself goes into the JSON. If the key lives in a Langflow global variable, only the variable name is exported, and the receiving instance needs a global variable with the same name. Use global variables from the start and exports stay safe to share (per the import and export docs, checked October 2026).

Troubleshooting the common first-flow errors

• Ports will not connect. The data types differ. Hover a port to see its type, or click it to search for compatible components, and add Type Convert where needed.
• "An API key must be passed as query or header". Your request has no Langflow API key. Add the x-api-key header, and check the variable is set in the same terminal session that runs the script.
• Model errors on the first run. The provider key lacks access to the model you picked, or the account has no credit. Enable the model in Model Providers and confirm billing with the provider.
• Windows Desktop install stops with a C++ error. Install Microsoft C++ Build Tools from the on-screen prompt, as the install docs describe.
• RAG answers ignore the document. The Load Data Flow did not run, or the two flows point at different collections or embedding models. Re-run the vector store component and compare the settings side by side.
• Every run repeats slow steps. Freeze the components whose output does not change, such as a finished ingestion step.

Where to go after your first flow

Once the API call works, you have the shape of a real AI feature: a flow that does one job well, behind an endpoint your app can reach. The usual next moves are giving the Agent more tools (many components can run in Tool Mode, and MCP servers can be tools too), swapping the vector store for a hosted one, and putting an automation tool in front so the flow runs when something happens in your business. For that last part, our Flowise vs n8n comparison helps you decide what sits where, and the guide to n8n AI workflows with OpenAI shows the trigger side. If retrieval is the core of your product, compare the managed route in our look at Vectorize as a RAG platform.

If you want to turn flows like this into a product people pay for, with auth, billing and a front end around the API call, our AI SaaS Builder program walks through that build end to end.

Langflow tutorial: FAQ