MMOOZZGG.com

How to build an AI agent

How to build a Flowise AI agent

8 min read14 sources

A Flowise AI agent is built with the mouse from blocks: a model, documents, tools and memory. The builder is released under the Apache 2.0 license, and the project has 55,000 stars on GitHub. Below are the installation, a first agent on an open model, a knowledge base and a website widget, and each step is checked against the Flowise documentation.

On this page · 5 sections

Build it yourself in an evening

Installation

You can install Flowise on your computer or server in two ways: with npm or with Docker. npm is the package manager for Node.js, the environment Flowise runs in. Docker runs programs in isolated containers.

1. Install Node.js version 18.15.0, or 20 and higher.

2. In the terminal, run the command npm install -g flowise.

3. Start the builder with the command npx flowise start, then in your browser open http://localhost:3000.

4. The Docker option: download the project with git clone https://github.com/FlowiseAI/Flowise, go to the docker folder, copy the .env.example file to .env and run docker compose up -d. The address in the browser is the same.

You can stop the Flowise containers with docker compose stop and start them again with the same command, docker compose up -d.

5. For an open model, install Ollama from ollama.com: there are versions for macOS, Windows and Linux. Then download a model with ollama pull and the model’s name from the Ollama catalog, for example ollama pull qwen3.

The Flowise documentation warns that your own server needs skills in installation, database backups and updates. A laptop is enough for an evening test, and work with customers needs a server that someone looks after.

A first agent from blocks

In Flowise an agent is built in a chatflow: a canvas where blocks are connected with lines. The agent block takes the model, tools and memory on separate inputs.

1. Open http://localhost:3000, click “Chatflows”, then “Add New”. Enter a name and click “Save”.

2. Click “Add Node”, find “Conversational Agent” and drag it onto the canvas. This is an agent for a dialog with a person.

3. Add the “ChatOllama” block from the “Chat Models” section. In the “Model Name” field, enter the model’s name from Ollama.

4. In the “Temperature” field, set a value from 0 to 1; the documentation suggests 0.5. Temperature sets how much answers vary: at a low value the agent answers evenly and to the point, at a high one more varied.

5. Connect the ChatOllama output to the agent’s “Chat Model” input.

6. Add memory from the “Memory” section and connect it to the “Memory” input. The “Buffer Memory” block stores the conversation in the chat_message table of the Flowise database, so it needs no separate server.

7. Connect tools to the “Allowed Tools” input. The access key for a service is set in the tool block with the “Connect Credentials” button, then “Create New”.

In the documentation example, the agent gets web search through the “SearchAPI” block, which needs a key for that service. For internal tasks, connect the knowledge base from the next section instead of search.

8. Click “Save Chatflow”, open the chat window with the chat icon and ask the agent a question.

For complex setups there is Agentflow V2. There, branches, pauses for a person’s reply and several agents are built as separate blocks. A chatflow is enough for a first agent, and a step-by-step plan for the first evening is on the page about a first agent in one evening.

Knowledge base

A knowledge base gives the agent your documents. Before it answers, the agent searches the base for the right pieces and answers from them, and the model does not need training. This approach is called RAG, and the page on how to build RAG covers it step by step.

1. Open the “Document Stores” section and add a store with a clear name.

2. Choose a document loader, for example “PDF Loader” for PDF files, and upload the files. You can add your own label to the loader, to filter the pieces of this set later.

3. Choose how to split the text, for example “Recursive Character Text Splitter”. Splitting cuts a document into pieces, so that the agent finds the right paragraph and not the whole file.

4. Look in the preview at how the pieces are cut and start the data processing. After processing you can edit or delete a piece by hand.

5. Click the “Select Embeddings” card and choose an embedding model. An embedding records the meaning of a piece as a set of numbers. For your own server there is the “Ollama Embeddings” block.

6. Click “Select Vector Store” and choose the vector database where these numbers will be stored. If you like, add a “Record Manager”: it stops the same piece from being loaded twice.

7. Click “Upsert” to load into the database, then “Retrieval Query” to test the search on a test question.

8. Connect the store to the agent. In a chatflow, this is the “Document Store (Vector)” block and the “Retriever Tool” on the “Allowed Tools” input. In Agentflow V2, you choose the store right in the “Agent” block and write in the “Describe Knowledge” field when to use it.

By default, search returns the 4 closest pieces; this is the top K parameter. More pieces give the model more context, but each answer costs more. How to choose a vector database is covered separately.

Adding it to your website

Flowise puts a finished agent on a website as a widget. A widget is a small script that draws a chat window on the page.

1. Copy the widget script that Flowise gives for the chatflow and paste it into the page code between the <body> and </body> tags.

2. Check two Chatbot.init parameters in the script: chatflowid, the ID of your chatflow, and apiHost, the address of your Flowise server.

To see which documents an answer relies on, pass the chatflowConfig parameter with the field returnSourceDocuments: true to Chatbot.init. The sessionId field there sets the conversation number.

3. Connect a domain and an HTTPS certificate to the server, as the Flowise documentation describes for deployment on your own server.

4. Turn on rate limiting: the “Message Limit per Duration” and “Duration in Seconds” fields. For example, 20 messages per 60 seconds from one IP address. An open chat without a limit is available to anyone who finds the page.

5. If the server is behind a proxy, set the NUMBER_OF_PROXIES variable, otherwise the limit may work incorrectly. The documentation suggests starting with 0, opening the /api/v1/ip address and adding 1 at a time until it shows your IP address.

If your Flowise is older than version 2.1.0, the embedded chat may not receive messages. The documentation suggests updating Flowise.

Visitors write their names, phone numbers and questions to the agent, and that is personal data. For a Russian company, 152-FZ, Russia’s personal data law, requires collecting and storing it in databases in Russia and notifying Roskomnadzor, the data protection regulator, before any transfer abroad. So for such an agent, Flowise, the model and the vector database run on a server in Russia; more on the page about data security.

How to plan an AI agent for a website that takes requests and hands the conversation to a person, we covered separately.

License

Flowise is released under Apache 2.0. The license allows you to copy the program, change it, distribute it and grant sublicenses free of charge, including in a commercial product.

When you distribute your own version, Apache 2.0 requires you to give the recipient the license text and keep the copyright notices.

There is an exception: the packages/server/src/enterprise folder and specially marked files come under a commercial license. These are enterprise features; the rest of the code stays on Apache 2.0.

In August 2025 Workday bought Flowise. On the day of our check, the license of the main code had not changed, but before you go live, open the LICENSE.md file in the repository.

A typical Flowise agent is built from open blocks. If someone charges for it as a unique development, ask what the price consists of.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. GitHub, 16 October 2026 (github.com)
  2. arXiv 2005.11401, 22 May 2020 (arxiv.org)
  3. Brace Law Firm, 9 July 2023 (brace-lf.com)
  4. Flowise Docs, 2026 (docs.flowiseai.com)
  5. Flowise Docs, 2026 (docs.flowiseai.com)
  6. Flowise Docs, 2026 (docs.flowiseai.com)
  7. Flowise Docs, 2026 (docs.flowiseai.com)
  8. Flowise Docs, 2026 (docs.flowiseai.com)
  9. GitHub, FlowiseAI/Flowise, 2026 (github.com)
  10. SiliconANGLE, 15 August 2025 (siliconangle.com)
  11. Flowise documentation: ChatOllama, checked 16 October 2026 (docs.flowiseai.com)
  12. Flowise documentation: Agentflow V2, checked 16 October 2026 (docs.flowiseai.com)
  13. Flowise documentation: Rate Limit, checked 16 October 2026 (docs.flowiseai.com)
  14. Ollama documentation: Quickstart and CLI Reference, checked 16 October 2026 (docs.ollama.com)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us