Open source AI agents:
what actually exists, and what "open" means
Anyone searching for this already ruled out the cloud: they want to see the code and run it themselves. This is the map of the ecosystem, with what each project solves and in what specific sense it is open — because the word covers four different things and usually only one of them is the one you care about.
First, the word
"Open" means four things and almost never all four at once
Before comparing projects it helps to know what you are asking for, because most arguments about "open source" agents are really two people talking about different dimensions.
The code licence. If it is OSI (MIT, Apache 2.0) you can use and modify it with no commercial conditions. Several popular projects use fair-code or "source-available" licences that publish the code but restrict commercial use: not the same thing, and if your company has a licensing policy that is the first filter.
Where it runs. An MIT-licensed agent can still send every prompt to a remote API. Open source does not imply local execution, and for most people searching for this, local execution is what they actually wanted.
Whose model it is. Almost all of these agents use third-party models over an API. Open weights is a separate question — Llama, Mistral, Qwen — and it is only answered at the model layer, not the agent layer.
The ecosystem
Six projects and what each one solves
None of these are ours. They are here because they are the real answers to this search, and a page that left them out would be worth nothing.
Aider
AI pair programming over your git repository, from the terminal. Very direct, very focused: it edits files and makes commits. If that is exactly what you want, it is hard to beat on simplicity.
OpenHands
An agent that attempts the full development loop: read the issue, write the code, run it and correct. It runs in a container and is among the most active projects on real issue-resolution benchmarks.
CrewAI
A framework for coordinating several agents with roles and divided tasks. Not an app you install and use: a library you build your own system on top of, in Python.
LangGraph
For when you need fine control over an agent's flow: states, cycles, checkpoints and a human in the loop. More ceremony than CrewAI, and in exchange far more control over what happens.
AutoGPT
The one that put the term into circulation in 2023. Today it matters more as a historical reference and a block-based platform than as the tool you would install to work, but it is still where a lot of people start.
Ollama, llama.cpp
Not agents: the way to run open-weight models on your own machine. If your real requirement is that no token leaves your network, the answer starts here and not at the agent layer.
The dimensions, separated
What each approach gives you, dimension by dimension
Check the current licence in each project's repository before deciding: they change, and more than one has.
| What you need | Where it is solved | Watch out for |
|---|---|---|
| Auditing what the agent does | Any project with published source, OSI or fair-code | Whether the published source is what the binary you installed actually runs |
| Commercial use with no conditions | OSI licences only: MIT, Apache 2.0, BSD | Fair-code and "source-available" do NOT meet this, even with the code in plain sight |
| The prompt not leaving your network | An open-weight model served locally (Ollama, llama.cpp, vLLM) | An MIT-licensed agent still calls a remote API if that is how you configure it |
| Files not leaving your machine | Local execution of the agent — independent of the licence | Agents that "run locally" but upload the context to a service to reason |
| Paying no subscription | Any self-hosted project, plus a local model if you want real zero cost | With an API model the bill does not disappear: it moves to the provider |
| Working without setting anything up | Desktop apps, including proprietary ones like Lienox | Here you usually give up the open licence: that is the real trade |
| Being able to change the behaviour | Frameworks like LangGraph or CrewAI, where you write the system | The cost is that you no longer install a tool: you maintain a project |
How to choose without losing the afternoon
Five questions, in this order
Is the requirement the licence, or privacy?
If it comes from an internal licensing policy, filter for OSI and drop the rest without opening them. If it comes from data not leaving, the licence is irrelevant and what matters is where it runs. Confusing these two is what costs people days.
Do you want a tool or a library?
Aider and OpenHands are installed and used. CrewAI and LangGraph are frameworks: they hand you pieces and you build. If you expected the first and downloaded the second, the disappointment is not the project's fault.
Is your work only code?
Almost the entire open ecosystem is built for development. If what you want to automate is research, content, ops or things that happen in a browser, the catalogue shrinks dramatically — and that is where proprietary software tends to show up.
How much maintenance will you accept?
A container, a database and monthly upgrades are free in licence and not in time. If nobody on the team is going to own that, a self-hosted project turns into debt within three months.
Can you try it on something real today?
Pick one concrete task that already annoys you and run it through two candidates. Feature comparisons — this one included — are for ruling out, not for choosing. Choosing happens by trying.
Where Lienox fits
And where it does not
Lienox is not open source. It is proprietary, with a permanent Free plan. If you got here because you need an OSI licence, that is the answer and we are not going to dance around it: the projects above are your candidates.
It does meet the other half of the search. Execution is local: your agents' prompts, files and results never leave your machine. The only thing that leaves is the inference call to the provider you choose with your own key — and if you point it at a local model, not even that. Telemetry is an allowlist of anonymous events: never content, paths or results.
What it adds over the open ecosystem is reach outside the repository: several agents coordinated on a board, a browser with a real cursor for anything that only exists as a web app, scheduled runs, and control by voice or Telegram. If your automation lives entirely inside git, you do not need it.
Frequently asked questions
Open source AI agents
What is the best open source AI agent?
There is no single best: there is one per kind of work. For editing code over a repository, Aider. For the full loop of resolving an issue, OpenHands. For building your own multi-agent system, CrewAI or LangGraph depending on how much control you need. And if what you want is that no token leaves your network, the decisive piece is not the agent but the local model you put underneath it.
Does open source mean my data stays on my machine?
No, and this is the most expensive confusion in the topic. The licence says what you may do with the code, not where your data travels. An MIT-licensed agent calling a remote API sends your prompts out exactly like a proprietary one. If your requirement is privacy, what you have to audit is where it runs and which endpoints it calls, not the LICENSE file.
Can I run an AI agent fully offline?
Yes, by combining an agent that executes locally with an open-weight model served on your own machine through Ollama, llama.cpp or vLLM. It works, and it is the only configuration with a real guarantee that nothing leaves. The price is quality: the models that fit on a desktop machine are still behind the best API models, and on long agentic tasks the gap shows.
Is n8n an open source AI agent?
It is a workflow automation tool with AI nodes, not an autonomous agent, and its licence is fair-code (the Sustainable Use License), which is not OSI-approved. It can be self-hosted and the source is published, but there are commercial-use restrictions. Both distinctions matter if you arrived searching strictly for "open source".
Is Lienox open source?
No. Lienox is proprietary, with a permanent Free plan and fully local execution. It is on this page because it covers half of what people usually want here — data not leaving your machine, without setting up infrastructure — and not the other half, which is the licence. If the licence is your hard requirement, stay with the projects above.
What is the difference between an agent framework and a ready-to-use agent?
A framework like CrewAI or LangGraph gives you the pieces to build a system: you write the code, define the roles and maintain the result. A ready-to-use agent like Aider or OpenHands installs and works. The confusion is expensive in both directions: pick a framework and you end up maintaining an internal project; pick a tool and you end up unable to change the thing you needed to change.
How much does running an open agent cost?
The licence, nothing. The rest depends. With an API model the bill moves to the provider and on long agentic tasks it climbs faster than almost anyone expects, because every step re-reads context. With a local model the cost is hardware and electricity. And in both cases you have to count the time to maintain the install, which is the line item that never appears in comparisons.
Local execution, without setting up infrastructure
Free forever on Linux, Windows and macOS. Your prompts and files never leave your machine — bring your own AI keys.