Which Open-Source AI Tools Are Actually Open Source?

Most self-hosted AI platforms marketed as open source are open core: a permissive licence on the bulk of the code, and a separate commercial licence on a carved-out directory. Of the seven largest tools in this category, five are not under an OSI-approved licence at all. GitHub already knows this and labels them `NOASSERTION`, which is its way of saying this is not a standard licence.

That is not automatically a scandal. Open core is a legitimate business model and these are genuinely useful products, as our Open WebUI review and Onyx review both show. The problem is narrower and more expensive: the features held back are the same three every single time. Single sign-on, role-based access control, and permission synchronisation from your source systems.

Those three are exactly what a European deployment needs to be lawful, as our EU AI governance guide sets out. We read every LICENSE file and every enterprise directory on 4 August 2026. Every figure below is checkable in about a minute.

A licence is open source in the strict sense only if it meets the Open Source Initiative definition: you may use, modify and redistribute the software for any purpose, including competing with the author. MIT and Apache 2.0 qualify. Almost everything else in this list does not.

The three restriction patterns we found:

1. Enterprise carve-out. Most of the repo is permissive, but a directory named `ee` or `enterprise` carries a commercial licence. Onyx, n8n and Flowise all do this, as our detailed Onyx review shows in one product.
2. Anti-competitive clause. You may use it, but not to operate a competing multi-tenant service. Dify does this.
3. Branding lock. You may not remove the vendor's name or logo above a certain size of deployment. Open WebUI and Dify both do this.

None of these is hidden. All of them are missed by nearly every comparison article, because reading a LICENSE file is less fun than copying a feature table.

This is the finding that matters for a procurement decision. Look at what is actually inside each enterprise directory:

- Onyx `backend/ee/` holds `external_permissions` (permission sync for Confluence, Google Drive, SharePoint, Slack, Salesforce, Jira and six more), `auth/users.py` for SSO, and `access/` for RBAC.
- Flowise `packages/server/src/enterprise` holds `rbac`, `sso`, `middleware` and the identity manager.
- n8n gates every file with `.ee.` in its name behind the n8n Enterprise License.

Three separate companies, three separate products, and the same three capabilities are behind the paywall in each. Single sign-on, role-based access control, and knowing who is allowed to see which document.

That is not a coincidence. Those are the features an enterprise cannot deploy without, which makes them the natural place to put the price.

We noticed something while reading: the Onyx Enterprise License and the n8n Enterprise License are near-identical, down to the sentence structure. Both permit copying and modifying for development and testing without a subscription, and both require a valid licence for the correct usage in production. Different companies, different countries, same boilerplate.

It tells you this is a settled industry pattern rather than any one vendor being sneaky. It also tells you what to search for: if a repo has a second licence file, the interesting part is which directory it applies to.

Art. 32 GDPR requires appropriate technical measures to ensure that only authorised people can access personal data. If you index a shared drive containing HR files with a free edition that has no permission synchronisation, your AI assistant will happily surface a salary letter to whoever asks the right question.

That is not a bug report. It is an unauthorised disclosure, and the fact that the software was free will not appear anywhere in the assessment. The same logic covers works council agreements: a system that cannot enforce role boundaries cannot honour the access rules you negotiated.

The practical consequence is simple. For any European deployment touching personal data, the free tier of an open-core AI tool is a proof of concept, not a plan. Budget for the enterprise licence from the start, or choose a tool whose permission layer is genuinely included. Our Glean alternatives guide compares both routes. See our guide to European AI data sovereignty and the detailed Onyx review for what this looks like in one specific product.

LibreChat and AnythingLLM are plain MIT with no enterprise carve-out, no branding clause and no anti-competitive restriction. We checked their repository roots for a second licence file and found none.

AnythingLLM goes further and publishes a self-hosted terms document stating that Mintplex Labs does not host, store or access any of your documents, chat history or embeddings, and that the software can run fully air-gapped. One honest caveat they document themselves: some default embedding models are fetched from their CDN on first use, so a truly air-gapped install needs those downloaded manually.

That is what an unrestricted licence looks like. It is worth noting that both are chat and assistant layers rather than full enterprise search platforms, so the comparison is not like for like on features. But if licence purity is a hard requirement in your organisation, this is the honest shortlist.

Open core is not dishonest. Every restriction we found is published in plain text in a file anyone can open, and these companies have to make money somewhere. The failure is in the ecosystem around them: comparison articles copy the word open source from a README and never check what it means in the specific repository.

So the rule is short. If a tool is going to hold your company's documents, spend the ten minutes. Look up the SPDX identifier, find the second licence file, list the enterprise directory. If SSO, RBAC or permission sync live in there, you are pricing the enterprise tier whether you planned to or not.

And if what you actually need is not document search but an understanding of how your organisation is functioning, none of these seven tools answers that question at any price.