You need an AI agent to sort your emails, answer quote requests, or prepare your reports, and you don't want to hand your customers' data to a cloud service whose server location and exact usage you don't know. That's where open-source frameworks come in: the code is public, you can read it, and you can run it on your own machines. You still need to know what these tools really do, what they cost, and what "open source" doesn't mean.
Key Takeaways
- Three frameworks dominate: LangGraph (control and enterprise workflows), CrewAI (simplicity), and Microsoft Agent Framework, the successor to AutoGen available since April 2026 under an MIT license.
- Gartner forecasts that 40% of enterprise applications will embed dedicated AI agents by the end of 2026, up from less than 5% in 2025.
- Open-source code doesn't mean free: you still pay for the model, either through a cloud API or through the hardware needed to run it locally.
- Open-source code doesn't mean no-code: these are developer tools. The no-code layer comes from a visual tool or a partner.
- The decisive criterion for an SMB isn't the framework, it's where the model runs.
Why open source concerns you
An AI agent framework is the toolbox that combines a language model, instructions, and external tools to execute a task. The open-source version publishes that code under a free license, most often MIT or Apache. Two consequences concern you, even if you don't write a line of code.
The first is control. No one can shut down your agent, change its terms of use, or raise its prices overnight. You depend on a community and your own infrastructure, not a vendor.
The second is confidentiality. Public code can be audited, and above all it can run on your machines. Your data stays with you. For independent professionals bound by professional secrecy, this is the point that changes everything. We detailed this architecture in our article on local AI and shared computing.
The landscape in 2026
Three frameworks stand out, along with a family of visual tools.
LangGraph: the choice for control
LangGraph, from the LangChain ecosystem, models an agent as a graph: each step is a node, each decision an edge. This structure gives fine-grained control over state, memory, and the points where a human must approve before the agent continues. It's the framework for long workflows and enterprise applications. It shows around 34,000 stars on GitHub as of early 2026. The downside: the steepest learning curve of the three.
CrewAI: the choice for simplicity
CrewAI organizes agents like a team: each one gets a role, a goal, and tools. Configuration goes through readable YAML files, which makes it accessible to a developer new to agents. It passed 50,000 GitHub stars in early 2026. The downside: past a certain level of complexity, the abstraction layers become opaque and debugging gets harder.
Microsoft Agent Framework: the successor to AutoGen
AutoGen, launched by Microsoft Research in 2023, was long the reference for multi-agent systems. In October 2025, Microsoft merged it with Semantic Kernel into a single framework, the Microsoft Agent Framework, available in a stable release since April 2026 under an MIT license, with a long-term support commitment. AutoGen remains in maintenance mode: security fixes only. If you're starting a new project, go with the new framework.
Visual tools: n8n and Dify
The three frameworks above require Python code. Two tools offer a visual approach. n8n chains steps together by drag-and-drop, without writing code (see our tutorial on building an agent with n8n). Dify adds a full interface for building assistants, with support for hundreds of models. These are the entry points when code isn't an option.
| Framework | Origin | Approach | Best for |
|---|---|---|---|
| LangGraph | LangChain | State graph | Enterprise workflows, fine-grained control |
| CrewAI | Independent | Roles and YAML | First multi-agent projects |
| Microsoft Agent Framework | Microsoft | Multi-agent, .NET and Python | Microsoft environments |
| n8n | Community | Visual no-code | Automation without a developer |
| Dify | Community | Visual interface | Custom assistants |
What "open source" doesn't mean
Two misunderstandings circulate, and they cost dearly for those who don't see them coming.
The first: open source doesn't mean free. The code is free, but the agent relies on a language model, and that's what gets billed. Two options: call a model via API and pay per token, or run a local model like Llama or Mistral on your own hardware, at the cost of a GPU-equipped server. Open-source models represent only about 11% of the model market, according to Menlo Ventures' annual report (2025). Most companies continue to rent proprietary models, despite the progress of open models.
The second: open source doesn't mean no-code. These frameworks are aimed at developers. An SMB that wants an agent without hiring a technical profile goes through a visual tool like n8n or Dify, or hands the build to a partner who knows these building blocks. That's the role of our custom AI agent service: assembling the right open-source components, then running the agent in the customer's environment. We also train teams to become self-sufficient with these tools (our automation training).
Confidentiality, the real reason
There's a more serious reason than price to choose open source: your data. A law firm, an accounting practice, or a medical practice cannot send client files to a cloud service without risk. With an open-source framework and a local model, nothing leaves the practice's infrastructure.
One nuance to watch for. Some development kits presented as open source (OpenAI Agents SDK, Claude Agent SDK, Google ADK) publish their code, but still send your requests to the vendor's API. The code is open, your data isn't. For a regulated profession, the real question isn't "is the framework open source?" but "where does the model run?"
Three questions before you choose
- What skill level do you have in-house? Without a developer, aim for n8n or Dify, or go through a vendor. With a technical team, choose LangGraph or CrewAI depending on complexity.
- Where does the data need to stay? If professional secrecy requires it, a local model is a must, whatever the framework.
- Who maintains the whole thing? An open-source framework evolves fast: versions follow one another, and a project can move into maintenance mode, as happened with AutoGen. Someone needs to track updates and fix breakages. An audit of your processes helps ground these choices in real use cases, not assumptions. To compare tools without a technical team, a hands-on workshop on AI agents sorts through the options using your own cases.
FAQ
What's the difference between an open-source framework and a kit like OpenAI's?
The framework orchestrates the agent (steps, memory, tools) and stays independent of the model provider. A development kit is often tied to one provider's API: the code is public, but requests go to its servers. Always check where the model runs.
Is an open-source framework enough for an SMB without a developer?
No. These tools require code. For an SMB without a technical team, the realistic path goes through a no-code tool like n8n or Dify, or through a partner who builds and maintains the agent.
Can you run an open-source agent on your own servers?
Yes, and that's the main appeal for sensitive data. You need a properly sized server and the skills to install and maintain it.
Is AutoGen dead?
Not dead, but in maintenance mode. Microsoft announced the merger with Semantic Kernel in October 2025, and the Microsoft Agent Framework has been stable since April 2026. Existing AutoGen projects keep running, but a new project should start with the new framework.
How much does an agent built on these frameworks cost?
The framework is free. The cost comes from the model (API or local hardware) and build time. For an SMB, the investment runs into the thousands of euros, but it varies a lot depending on the complexity of the use case.
Conclusion
Choosing an open-source framework comes down to answering three questions: the level of control you want, where your data needs to stay, and the skills you have. LangGraph for control, CrewAI for simplicity, the Microsoft Agent Framework for Microsoft environments, and visual tools when code isn't an option.
For a French SMB or independent professional, open source isn't a trend: it's often the only way to use an AI agent without giving up control of your data. The framework is just one component. What matters is the whole picture: a model running in the right place, tools connected properly, and maintenance that keeps up.
That's the work NexeAI does for its clients. If you want an AI agent built on open-source building blocks, hosted where your data requires it, let's talk about your project.
