Open-source AI agents you can self-host (and what it really costs)
Seven open-source and fair-code agent projects compared on the exact licence, what you need to run them, a EUR 6 server cost model and which repositories were archived this year. Licences and prices read on 2026-11-09; our own deployment log is not yet published, and we say so.
Independent editorial content. Links to software go straight to the vendors. We are not in any affiliate programme yet and earn nothing from these links; if that changes, this note will say so. Rankings are never sold. See our methodology. Methodology
| Tool | Licence (read 2026-11-09) | Self-host requirements stated | Interface | Hosted option (checked 2026-11-09) | Who should not pick it | |
|---|---|---|---|---|---|---|
| n8n Visual canvas, fair-code, EU cloud if you want it | Sustainable Use License 1.0; .ee files excluded; internal business or personal use only | Docker; no CPU/RAM figure | Visual canvas plus code nodes | Cloud Starter EUR 20/mo annual (2,500 executions), Pro EUR 50; Frankfurt | Agencies reselling it as a hosted service | Visit |
| Dify Chat apps and knowledge bases with a near-Apache licence | Dify Open Source License: Apache 2.0 plus no multi-tenant use without authorisation and no logo removal | CPU ≥ 2 cores, RAM ≥ 4 GiB, Docker Compose ≥ 2.35 | Visual canvas, chat apps, knowledge base, API | Sandbox free; Professional USD 590/workspace/year; Team USD 1,590/year | Multi-client hosting without a written licence | |
| Flowise Archived August 2026 | Apache 2.0 outside enterprise directories (commercial licence inside) | Node.js ≥ 20; Docker | Drag-and-drop | Homepage listed Free, Starter USD 35/mo, Pro USD 65/mo while announcing a sunset | Anyone starting a new deployment | |
| CrewAI MIT multi-agent framework for Python developers | MIT; no exclusions stated | Python ≥ 3.10 and < 3.14 | Code framework and CLI | AMP Suite and Control Plane described as free to try; no prices shown | Non-developers | |
| LangGraph MIT low-level orchestration with durable state | MIT; no exclusions stated | pip install; nothing else stated | Code framework; visual prototyping in separate LangSmith Studio | LangSmith Deployment linked, no prices shown | Teams wanting a canvas their colleagues can edit | |
| AutoGen Archived; Microsoft points to Agent Framework | MIT code, CC-BY-4.0 docs | Python ≥ 3.10 | Code framework; Studio GUI not production-ready per README | None | New projects | |
| OpenHands (Agent Canvas) MIT control centre for coding agents | MIT; no exclusions stated | Node.js ≥ 24 and uv; Docker for sandbox options | Web control centre for coding agents | OpenHands Cloud and Enterprise offered; no prices shown | Anyone running it without the Docker sandbox |
Prices and features checked: 2026-11-09. Prices in the currency the vendor shows, checked on the date shown; we do not convert.
Verdict by user type
"Open-source AI agent" covers two different things: a visual platform you run on a server and hand to colleagues (n8n, Dify, Flowise), and a code framework a developer imports (CrewAI, LangGraph, AutoGen, OpenHands). The licence, the hosting bill and the maintenance hours differ sharply between the two, and so does who should run them.
Our criteria, stated first: the exact licence and what it excludes, read from the repository on 2026-11-09; what the project requires to self-host (Docker, Node, Python, minimum RAM where stated); the interface a non-developer can use; which model providers are supported; whether the project is still maintained; and what a small EU server costs from a published price list. n8n is the only product below with an affiliate programme (in our catalogue); its position does not depend on that.
A disclosure that matters for this page in particular: we have not yet run and logged our own instance of each project. Our plan called for a deployment log with monthly cost and maintenance hours, and that log does not exist yet. The cost model below is computed from published server prices and the projects' stated requirements, and the maintenance estimates are reasoning, not measurement. We mark them as such.
| You are… | Pick | Why |
|---|---|---|
| A team that wants colleagues to build agents on a canvas, hosted in the EU | n8n (self-hosted or Cloud) | Visual canvas plus code, Docker install, Frankfurt hosting if you take Cloud; licence permits internal business use |
| A team that wants a chat-app and knowledge-base builder with a permissive-ish licence | Dify | Apache 2.0 base with two extra conditions; 2 cores and 4 GiB stated; cloud tiers if you change your mind |
| A Python developer orchestrating several specialised agents | CrewAI | MIT; Python 3.10–3.13; defaults to OpenAI, local models via Ollama |
| A developer who needs durable, inspectable state across long agent runs | LangGraph | MIT; low-level graph orchestration; visual tooling only in the separate LangSmith product |
| A developer who wants a coding agent that touches a real repository in a sandbox | OpenHands (Agent Canvas) | MIT; Docker sandbox optional; works with Claude Code, Codex and any ACP agent |
| Someone evaluating Flowise or AutoGen today | Neither | Both repositories are archived; see the maintenance section |
Why self-host at all: privacy, cost, control
Three reasons come up in every conversation, and only two of them survive contact with a price list.
Privacy and the GDPR. When an agent processes personal data, a hosted vendor is a processor under Article 28 of the GDPR and the controller must be able to document the instructions, sub-processors and security measures (Article 32) [11]. Self-hosting does not remove those duties; it moves them to you. What it does remove is the question "where is the vendor's server", because the answer is "ours, in Falkenstein". Search demand reflects the concern: "private ai agents" is searched about 5,400 times a month in the US and "self-hosted ai agents" 3,600, the latter with a keyword difficulty of 0 (DataForSEO, retrieved 2026-11-09).
Cost. This one survives only at volume. A shared-vCPU server at Hetzner starts at EUR 5.99 a month for the cost-optimised line (listed "currently not available" in part on the day we read it) or EUR 6.49 for regular performance, excluding VAT, with data centres in Falkenstein, Nuremberg and Helsinki [8]. Against n8n Cloud Starter at EUR 20 or Pro at EUR 50 [9], the server wins on paper from month one. It loses as soon as you count the hours, which the cost section below does.
Control. Which model, which version, which plugins, when to upgrade. For code frameworks this is the whole point; for visual platforms it is a mixed blessing, because every upgrade is yours to test.
The seven projects: licence, requirements, interface
| Project | Licence (read 2026-11-09) | Exclusions and conditions | Self-host requirements stated | Interface | Status |
|---|---|---|---|---|---|
| n8n | Sustainable Use License 1.0 ("fair-code") | Files with ".ee." in the name are not licensed; internal business or personal use only | Docker; no CPU or RAM figure | Visual canvas plus code nodes | Active; Series C at USD 2.5 bn valuation Oct 2025 |
| Dify | Dify Open Source License (Apache 2.0 plus conditions) | No multi-tenant operation without written authorisation; front-end logo and copyright may not be removed | CPU ≥ 2 cores, RAM ≥ 4 GiB, Docker Compose ≥ 2.35 | Visual canvas, chat apps, knowledge base, API | Active |
| Flowise | Apache 2.0 outside enterprise directories | packages/server/src/enterprise and marked files under a commercial licence | Node.js ≥ 20; Docker | Drag-and-drop | Archived 13 Aug 2026; homepage says "We're sunsetting Flowise" |
| CrewAI | MIT | None stated | Python ≥ 3.10 and < 3.14 | Code framework and CLI | Active |
| LangGraph | MIT | None stated | pip install; nothing else stated | Code framework; visual prototyping in separate LangSmith Studio | Active |
| AutoGen | MIT (code), CC-BY-4.0 (docs) | None stated | Python ≥ 3.10 | Code framework; AutoGen Studio GUI "not meant to be production-ready" | Archived; README points new users to Microsoft Agent Framework |
| OpenHands (Agent Canvas) | MIT | None stated | Node.js ≥ 24 and uv; Docker for the sandbox options | Web control centre for coding agents | Active; renamed Agent Canvas |
Sources: the repositories and licence files [1][2][3][4][5][6][7].
On "fair-code" versus open source. n8n's licence lets you "use or modify the software only for your own internal business purposes or for non-commercial or personal use" and excludes the enterprise ".ee" files [1]. Running agents for your own company is inside it; selling n8n hosting to others is not. Dify's licence is Apache 2.0 with two conditions, the important one being that you "may not use the Dify source code to operate a multi-tenant environment" without written authorisation, where one tenant equals one workspace [2]. For a single company with a single workspace, that is no constraint. For an agency hosting Dify for several clients, it is. CrewAI, LangGraph and OpenHands are MIT, which imposes neither limit [4][5][7]. If your lawyer asks for "open source" in the OSI sense, n8n and Dify are not it, and the difference is worth one paragraph in your vendor file rather than a veto.
On maintenance status. Flowise's repository reads "archived by the owner on Aug 13, 2026. It is now read-only", and its homepage announces a sunset [3]. AutoGen's README opens with "Project archived and no longer maintained" and sends new users to Microsoft Agent Framework [6]. Both still run, and both will stop getting security patches. We list them so you can recognise the names in older tutorials, not so you deploy them.
What it costs on a EUR 6 server: a model, not a log
The arithmetic below uses Hetzner's published shared-vCPU prices [8] and the projects' stated requirements. It is a model. When our own deployment log exists we will replace it with measurements.
| Line item | n8n self-hosted | Dify self-hosted | Code framework (CrewAI, LangGraph) | Hosted comparison |
|---|---|---|---|---|
| Server | Regular-performance shared vCPU from EUR 6.49/mo excl. VAT; n8n states no minimum | Dify states ≥ 2 cores and 4 GiB, so expect a step up from the entry card; "see pricing" for the exact tier | Runs where your code runs; often no dedicated server | n8n Cloud Starter EUR 20, Pro EUR 50; Dify Professional USD 590 per workspace per year (about USD 49/mo), Team USD 1,590 per year [14] |
| Model calls | Billed by the provider; n8n Cloud includes Assistant credits, self-hosted is BYOK | Billed by the provider; Dify cloud includes 200 (Sandbox), 5,000 (Pro) or 10,000 (Team) message credits a month, then your own key | Billed by the provider | Included credits, then provider |
| Domain and TLS | A few EUR a year; a reverse proxy you configure | Same | Not needed for a script | Included |
| Backups | Your job; a snapshot or volume costs extra | Your job | Your repository | Included |
| Updates | Your job: pull, restart, test | Your job: Docker Compose pull, migrations | Your job: dependency bumps | Included |
Two conclusions hold regardless of the hours you assume. First, the server is not the cost; the model calls are, and they are identical whether you self-host or not. Second, the hosted price buys you updates, backups and someone else's on-call. A reasonable rule, the same one we use for automation platforms generally: self-hosting pays once the hosted plan you would otherwise need costs more than roughly EUR 50 a month, or when the residency requirement is non-negotiable and the vendor does not state a region. For n8n the residency argument is weaker than usual, because its Cloud plans already state Frankfurt [9].
On the hours: we will not print a number we did not measure. The honest range from the projects' own documentation is "one evening to install, an hour or two a month to update and check logs" for Docker-based platforms, and "whatever your developer's time costs" for frameworks, where the agent is part of your software rather than a separate service.
Security checklist for a self-hosted agent
An agent on your server has whatever your server has. Before you give it a key:
- Sandbox the actions. OpenHands warns that its no-sandbox option gives the agent "full access to your filesystem" and offers Docker sandboxes instead [7]. Use them. For n8n and Dify, run the container as a non-root user with a volume only for its own data.
- Scope the credentials. One API key per agent, with the narrowest scope the provider offers. Rotate when the agent's job changes.
- Put a human on the irreversible step. n8n ships "human approval for tool calls" as a feature [9]; in a framework you write that gate yourself. The automation-bias literature is unambiguous that people under-check automated aids unless accountability is designed in A[12], and a systematic review of 74 studies in clinical decision support found users followed wrong advice at non-trivial rates, with training and accountability as the mitigators A[13].
- Log every tool call with its inputs, and keep the logs where the agent cannot delete them.
- Document the processing. If personal data flows through the agent, write the Article 30 record and the Article 32 measures now [11]; self-hosting makes you the processor and the controller at once.
- Patch on a schedule. The archived projects above are the warning: an agent platform that no longer receives security fixes is a liability, not a saving.
When hosted wins
- You have no one who has run a Docker container, a reverse proxy and a backup. n8n Cloud Starter at EUR 20 [9] is cheaper than the first weekend.
- You need an SLA, SSO or audit logs you can show an auditor. Those are enterprise features on both sides, and the hosted vendor has already written the documentation.
- Your volume is low. Below a few thousand runs a month, the EUR 13 you save on hosting does not cover one hour of anyone's time.
- You want someone else to decide when to upgrade the model integrations. Provider APIs change; a hosted platform absorbs that.
For agents you subscribe to rather than run, see the best AI agents for work; for building agents without code, our no-code agent builders comparison costs out Zapier, Make, n8n, Lindy, Dify and Voiceflow at three run volumes; and for the open-source coding agents in particular (OpenHands, Aider, Codex CLI), see AI coding agents for non-developers.
The demand picture
"Open-source ai agents" is searched about 9,900 times a month in the US, up 233 percent year on year, with a keyword difficulty of 19; "open source ai agents framework" 1,600; "free open source ai agents" 1,000 (DataForSEO, retrieved 2026-11-09). In our Future Demand Radar, the "AI agents and delegated work" category scores 59.7 out of 100 in the baseline scenario (band 55.2–64.1) and is classed "build now". Our baseline scenario expects self-hosting interest to track the wider agent category, with the specific risk that the number of maintained projects shrinks as it did for Flowise and AutoGen this year; that is a forecast, not a fact. n8n's own trajectory, a USD 180 million round at a USD 2.5 billion valuation in October 2025 B[10], is the clearest signal that the visual, self-hostable end of the market is where the money has gone.
What to do next
If you have a developer and a residency requirement, install n8n or Dify on a EUR 6.49 server this week, wire one agent to one tool with a human approval step, and log everything for a month before you add a second. If you do not have that developer, take n8n Cloud and keep the option to move later; the licence lets you. Either way, run the Tool Finder with your run volume and residency requirement to see the same dated prices side by side, read how to manage AI agents before the first one goes live, and subscribe to the briefing below to get the self-hosting checklist as a one-page PDF when we publish our deployment log.
Frequently asked questions
Is n8n open source?
Not in the OSI sense. n8n is published under the Sustainable Use License 1.0 and describes itself as fair-code: you may use or modify it for your own internal business or personal use, you may not distribute it commercially, and files with ".ee." in the name are excluded. Running agents for your own company on your own server is inside the licence; reselling n8n hosting is not. Read 2026-11-09.
What is the cheapest way to self-host an AI agent in the EU?
A shared-vCPU cloud server at Hetzner starts at EUR 5.99 to 6.49 a month excluding VAT with data centres in Germany and Finland (checked 2026-11-09). n8n states no minimum hardware; Dify states at least 2 cores and 4 GiB of RAM. The server is rarely the real cost: model API calls and your maintenance hours are, and those are the same whether you self-host or not.
Which open-source AI agent frameworks are still maintained?
On 2026-11-09, n8n, Dify, CrewAI, LangGraph and OpenHands (now Agent Canvas) were active. Flowise's repository was archived on 13 August 2026 and its homepage announces a sunset; AutoGen's README states the project is archived and no longer maintained and points new users to Microsoft Agent Framework. Do not start new deployments on archived projects.
Does self-hosting an AI agent make me GDPR compliant?
No. It removes the question of where the vendor's servers are, but you remain the controller and become your own processor: Article 28 processor duties and Article 32 security measures still apply, and you need to document them. A hosted vendor that states its region, such as n8n Cloud in Frankfurt, can be the simpler path for a small team.
Is Dify free to use commercially?
Yes for a single company and a single workspace: the Dify Open Source License is Apache 2.0 with two extra conditions, no multi-tenant operation without written authorisation and no removal of the front-end logo and copyright. Hosting Dify for several clients requires permission. The cloud plans were Sandbox free, Professional USD 590 per workspace per year and Team USD 1,590 per year on 2026-11-09.
Sources and evidence
- n8n repository and Sustainable Use License (LICENSE.md) — n8n GmbH (2026), vendor page A github.com/n8n-io/n8n/blob/master/LICENSE.md (retrieved 2026-11-09)
- Dify repository, README and Dify Open Source License — LangGenius (2026), vendor page A github.com/langgenius/dify/blob/main/LICENSE (retrieved 2026-11-09)
- Flowise repository (archived 13 Aug 2026) and LICENSE.md — FlowiseAI (2026), vendor page A github.com/FlowiseAI/Flowise (retrieved 2026-11-09)
- crewAI repository, README and MIT LICENSE — crewAI, Inc. (2026), vendor page A github.com/crewAIInc/crewAI (retrieved 2026-11-09)
- LangGraph repository, README and MIT LICENSE — LangChain, Inc. (2026), vendor page A github.com/langchain-ai/langgraph (retrieved 2026-11-09)
- AutoGen repository (archived) and LICENSE — Microsoft (2026), vendor page A github.com/microsoft/autogen (retrieved 2026-11-09)
- OpenHands / Agent Canvas repository, README and MIT LICENSE — OpenHands contributors (2026), vendor page A github.com/All-Hands-AI/OpenHands (retrieved 2026-11-09)
- Hetzner Cloud: server prices and locations — Hetzner Online GmbH (2026), vendor page A www.hetzner.com/cloud/ (retrieved 2026-11-09)
- n8n pricing (Cloud plans, human approval for tool calls, Frankfurt hosting) — n8n GmbH (2026), vendor page A n8n.io/pricing/ (retrieved 2026-11-09)
- Hyped AI business automation startup n8n raises $180m, valuing it at $2.5bn — Tech.eu (2025), report B tech.eu/2025/10/09/hyped-ai-business-automation-startup-n8n-raises-180 (retrieved 2026-11-09)
- Regulation (EU) 2016/679 (General Data Protection Regulation), Articles 28, 30 and 32 — European Parliament and Council (2016), report A eur-lex.europa.eu/eli/reg/2016/679/oj (retrieved 2026-11-09)
- Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman R, Manzey DH (2010), systematic review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)
- Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard K, Roudsari A, Wyatt JC (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)
- Dify pricing (Sandbox, Professional, Team, Community, Enterprise) — LangGenius (2026), vendor page A dify.ai/pricing (retrieved 2026-11-09)