Your own company ChatGPT, on your own servers
Many companies are in the same place: employees already use AI chats, often with private accounts and without rules. Management wants to bring this under control. A company plan with a US service is booked quickly, but then chat histories and uploaded documents sit with the vendor.
This is where Open WebUI comes in. It is one of the best-known tools of its kind, and the questions behind it are always the same: is it free? Is it secure? Is it GDPR compliant? And are there alternatives?
I answer these questions from a business perspective. You won't find an installation guide here, there are plenty of those. This is about the decision: whether Open WebUI fits your company and what running it demands.
What Open WebUI is and who is behind it
Open WebUI is a chat interface for language models that you run on your own server. It looks and feels like ChatGPT. The difference: you decide which model answers and where the data lives.
The project started as «Ollama WebUI», an interface for locally running models. Today the company behind it is Open WebUI, Inc., based in San Francisco and founded by Tim J. Baek. The source code is public on GitHub and developed further by a large community.
Technically, Open WebUI connects to two kinds of model sources: Ollama for models on your own hardware, and any API that speaks the OpenAI format. Besides OpenAI itself, that includes European vendors such as Mistral or the IONOS AI Model Hub. The interface is replaceable, and so is the model behind it.
Is Open WebUI free? The licence in detail
Short answer: yes, with one condition. Since version 0.6.6 from April 2025, the project's own «Open WebUI License» applies. It builds on the permissive BSD-3 licence and adds a branding clause.
The clause says: if more than 50 people use the installation within 30 days, the Open WebUI name and logo must not be removed or replaced. If you still want the interface in your own design, you need written permission or an enterprise licence. The vendor does not publish its price.
For most companies that is no obstacle. An internal tool does not have to look like your house brand. Another consequence matters more: the Open WebUI License is not approved by the Open Source Initiative. If your legal department only allows approved open source licences, Open WebUI is out.
What applies to older versions
Code up to and including version 0.6.5 remains under the plain BSD-3 licence. In theory, you could stay on an old version or use a fork of it. In practice, we advise against it: you give up every security update since spring 2025. Why that is dangerous is shown in the security section below.
What Open WebUI can do for a business and where it ends
Four features matter for business use:
- Login with company accounts: Open WebUI supports OIDC, LDAP and SCIM. Employees sign in with their Microsoft, Google or Keycloak account, and groups are taken over from the directory.
- Roles and groups: Besides admin and user, there is a pending role for new accounts. Groups unlock additional permissions, for example for certain models or tools.
- Document search: Under «Knowledge», you upload documents the model draws on when answering. By default Open WebUI stores the vectors in Chroma; for production, PostgreSQL with pgvector can be configured.
- Multiple models: You can offer several models side by side and enable them per group, for example an inexpensive one for everyday questions and a strong one for analysis.
What does that look like day to day? Sales has a quote drafted from bullet points. Accounting summarises a twenty-page framework contract in five points. Service asks the uploaded manuals about an error code. All three work in the same interface, but with different models and documents.
The limit is permissions. Open WebUI grants access per knowledge collection and group. It does not take over the permissions of individual documents from SharePoint or the file server. For one department with shared documents that is enough, as it is for a service team where everyone may use the same manuals.
For a company where personnel files sit next to sales documents, it is not. Then you need your own knowledge base behind it that checks permissions per passage. How that is built is shown in our article on the AI knowledge base.
Is Open WebUI secure?
Open WebUI is as secure as the way you run it. A case from November 2025 shows this: the vulnerability CVE-2025-64496 affected all versions up to 0.6.34. Via a manipulated external model connection, accounts could be taken over and, in some cases, code executed on the server. Depending on the source, its severity was rated 7.3 to 8 out of 10.
It could only be exploited if the «Direct Connections» feature was enabled and someone had added a malicious model. It has been fixed since version 0.6.35. Anyone who installed Open WebUI once and then forgot about it is still running this vulnerability today.
That leads to three rules for operation:
- Fixed update routine: New versions of Open WebUI appear at short intervals. Plan at least monthly updates and apply security updates immediately.
- Enable features sparingly: Turn on external model connections and custom tools only for the groups that need them.
- No open access from the internet: Login only via company accounts, ideally also behind a VPN or an upstream access gateway.
In practice, a second installation as a test environment proves its worth. Every update runs there first, and a short check covers login, models and knowledge collections. Only then does the update go into production. That costs half an hour per update and saves you the Monday morning when nobody can get into the system.
The point: a self-hosted AI interface is not a device you set up, but a service you maintain.
Is Open WebUI GDPR compliant?
The question is framed wrongly, and that is good news. Open WebUI is software you run yourself. Whether it is GDPR compliant depends on how you run it. Four points decide:
- Server location: Open WebUI runs on a server in the EU, with a host that signs a data processing agreement.
- Model vendor: Every question goes to the model behind it. Choose a vendor that processes in the EU under an Article 28 GDPR agreement, or run the model yourself.
- Chat histories: Open WebUI stores all conversations in its database. Decide how long they are kept and who may see them.
- Rules for employees: A short usage policy defines which data may go into the chat. Where a works council exists, involve it in the rollout.
With these four points, Open WebUI is often better positioned on data protection than a company account with a US service. For more background, see our guide on GDPR-compliant AI. For your specific case, ask your data protection officer.
What running it costs
The software itself costs nothing. Costs arise for servers, model usage and maintenance. Three typical setups, prices net:
| Setup | Server | Model | Running costs per month |
|---|---|---|---|
| Interface plus EU API | cloud server, from €8.49 | usage-based billing | around €75 for 50 people |
| Interface plus own model | GPU server, from €232.30 | 14B-class model on your own hardware | from around €232 |
| Large own model | GPU server, €1,197.30 | 70B-class model | from around €1,200 |
Server prices are from Hetzner (GPU servers as of August 2026, cloud server as of October 2026). Model costs for the first setup follow our sample calculation: 50 people, 20 questions a day each, 20 working days, around 5,000 tokens per question. At €0.65 per million tokens for Llama 3.3 70B at IONOS, that is €65 a month.
What no table shows: maintenance. Updates, backups, monitoring and the connection to the company login take time, even when everything runs. Expect a few hours a month, in-house or with a service provider. The full calculation for your own models is in our article on LLM self-hosting costs.
Run it yourself or have it managed?
Running it yourself pays off if your IT already maintains Linux servers and containers. Then Open WebUI is one more service in existing operations. Without that routine, a managed setup is the more honest choice. The question is not who installs it, but who applies the update on the same day when the next vulnerability appears.
How to roll out Open WebUI in your company
Technically, a company chat is set up in a day. To make sure it is used and creates no new risks, we work in four steps:
- Pilot group: Ten to twenty people from two or three departments test it for four weeks. They bring real tasks instead of toy questions.
- Usage policy: One page is enough. It defines which data may go into the chat, how long histories are kept and whom to ask.
- Login and groups: Login via company accounts, groups from the directory, models and knowledge collections enabled per group.
- Review: After four weeks, review: which tasks work, where are documents missing, which model is enough? Only then does the rest of the workforce follow.
A side effect many underestimate: an official company chat brings private AI use out of the grey zone. People who get a good tool with clear rules stop copying customer data into private accounts. Bans alone don't stop shadow AI, a better offer does.
Open WebUI, LibreChat or AnythingLLM?
Two alternatives appear in almost every comparison. Both are MIT-licensed, without a branding clause.
| Criterion | Open WebUI | LibreChat | AnythingLLM |
|---|---|---|---|
| Licence | own licence with branding clause above 50 users | MIT | MIT (core) |
| Strength | quick to set up, large community | many vendors in parallel, own design allowed | documents at the centre, also as desktop app |
| Operation | lightweight | more components, more effort | server or single workstation |
| Fits | companies without a need for their own design | companies with their own design or strict licence review | small teams focused on documents |
Our assessment: Open WebUI is the pragmatic start for most mid-sized companies. Choose LibreChat if the licence question or your own design matters. AnythingLLM is strong when a small team mainly works with documents and has no IT of its own.
The good news: the choice is not final. All three use the same model APIs, so models and vendors stay the same. What can be lost when switching are old chat histories and the settings of knowledge collections. If you keep your documents in your own knowledge base, you can switch the interface later with little effort.
Our recommendation by starting point
Here is our position, as we present it in initial calls:
| Starting point | Our recommendation |
|---|---|
| First steps, up to 50 people | Open WebUI on an EU server, model via an EU API, login with company accounts |
| More than 50 people, own design wanted | LibreChat or an Open WebUI enterprise licence, ask for the price first |
| Legal requires an approved open source licence | LibreChat or AnythingLLM |
| Fine-grained permissions per document needed | Open WebUI as interface, your own knowledge base with permission checks behind it |
| Data must not leave the building | Open WebUI with your own model on a GPU server in the EU |
What we explicitly advise against
- Install and forget: Open WebUI without an update plan becomes a security risk within a few months, see CVE-2025-64496.
- Local accounts instead of company login: Local user accounts with their own passwords are easily forgotten when someone leaves. Login belongs to the central directory, for example via Keycloak or Microsoft Entra ID.
- Loading all company documents into one collection: Putting everything into one shared knowledge collection gives everyone access to everything.
Our position in one sentence: Open WebUI is a good starting point, as long as you run it like business software and not like a weekend project.
Next steps
If you want to introduce Open WebUI or an alternative, we clarify three questions together: which interface fits your licence requirements and team size? Which model and location fit your data? And who takes care of updates and maintenance?
In 30 minutes you will know how your company chat should be set up and what it costs per month. Book a no-obligation video call here. For the bigger picture, read our guide to adopting AI in your company.
