No-code builders
A builder assembles an agent from blocks on a diagram: input, model, tools, answer. You do not need to write code, and a simple agent takes an evening or a week to build.
n8n. An automation platform with AI agents, based in Berlin, working since 2019. Its fair-code license (Sustainable Use License) is not an OSI open source license: you may install it on your own server and change it, but you may not resell it as a cloud service. It fits when the agent is built into a process with spreadsheets, a CRM, and messengers.
Flowise. A visual builder for AI agents, Apache 2.0 license, 55 thousand stars on GitHub. GitHub stars show how many developers marked a project, a rough measure of its popularity.
Dify. A platform for AI applications and agents on Apache 2.0 with two conditions: without permission, you may not run a multi-tenant cloud service like Dify or remove the logo from the console. For an agent inside your own company, these conditions usually do not get in the way.
Each builder has step-by-step builds; the easiest start is the n8n guide.
Libraries for developers
A library is a set of ready code from which a programmer builds an agent. It gives more flexibility than a builder, but you need a person who writes Python or JavaScript.
LangChain. An open library for building agents and chains in Python and JavaScript, MIT license. A base for an agent with tools and document search; the build is covered in the LangChain guide.
LangGraph. A library from the same team, also MIT. In it, the agent’s work is described as a graph of steps, which helps when you need checks and a person in the loop.
CrewAI. A framework, that is, a ready program skeleton, for teams of agents with roles, tasks, and processes. MIT license, first release in December 2023.
OpenAI Agents SDK. An open kit released on 11 March 2025: agents, handoffs between them, guardrails, and tracing. MIT license; besides the OpenAI API, it works with more than 100 other models.
Claude Agent SDK. Since 29 September 2025, this is the name of the former Claude Code SDK: a loop of “gather context, take action, check the result, repeat” and services connected through MCP. The code is marked with the MIT license, but use is governed by Anthropic’s commercial terms, and Claude Code is built into the package.
Models and running them
An agent needs a model. You download an open model and run it on your own machine, and then requests do not go to its developer.
Ollama. A program that downloads open models and runs them on a computer or server, MIT license. A local agent starts with it.
DeepSeek R1. The code and the weights of the model, that is, the trained model itself as files, are open under the MIT license. The developer directly allows commercial use, changes, and derived models.
Qwen 3. A family of open-weight models released on 29 April 2025 under the Apache 2.0 license. There are standard models from 0.6 to 32 billion parameters and two large mixture-of-experts models with 30 and 235 billion.
Which other models suit an agent and how to compare them is covered in the section on open models.
The MCP connection standard
MCP (Model Context Protocol) is a single way to connect AI to data and services. Anthropic opened the protocol on 25 November 2024, and the specification is on GitHub under the MIT license.
MCP works like a standard power socket. A service builds an MCP server once, an adapter program for its data, and any agent that understands the protocol can connect to it.
For a business, this means less work to connect a CRM, email, or a document base: you do not need a separate integration for each model. The Claude Agent SDK, for example, connects outside services through MCP.
Before you connect someone else’s MCP server, check which data it will get access to and who wrote it. A server with write access to the CRM is more dangerous than one that only reads.
How to choose
The choice starts with the task, not the tool. First describe the process, the data, and the metric, then pick the blocks.
The agent is built into a process with spreadsheets, a CRM, and messengers: n8n. The agent answers from documents: Flowise or Dify. A team of agents with roles: CrewAI. Your own logic and checks: LangChain and LangGraph. Data must not leave the company: Ollama and an open model.
Check the license before launch. n8n may not be resold as a cloud service, Dify has two conditions, the Claude Agent SDK works under Anthropic’s terms, and the other tools on the list are on MIT or Apache 2.0.
A typical agent is built from these open blocks. If you are asked to pay for it as for unique development, ask what the price is made of, and pay for the result by the metric, not for hours.
Which agents you can build from these tools with no license costs is covered on the page on free agents.