Onyx AI Review 2026: What 31,400 GitHub Stars Actually Buy You

Onyx (formerly Danswer) is an open-source AI search and chat platform that connects large language models to your company's internal documents, tickets and wikis. You can self-host it for free under an MIT licence, or buy Onyx Cloud at 20 US-Dollar per user per month. It is the most credible open-source answer to Glean, with 31,400 GitHub stars and production users including Netflix, Ramp and Thales Group.

That is the part every comparison article gets right. Here is the part all of them get wrong: the permission layer is not MIT-licensed. The code that makes Onyx respect who is allowed to see which document lives under a directory called `ee`, and that directory carries a separate commercial licence requiring a paid Enterprise seat count for production use.

We did not take anyone's word for this. We cloned the repository on 4 August 2026 and read the licence files, the enterprise directory and the deployment manifests. Every claim below is checkable against the public repo.

Onyx is a self-hostable AI workspace: it indexes your company data through 51 connectors, then lets employees ask questions in natural language and get answers with inline citations back to the source document. It works with any model provider, from OpenAI and Anthropic to fully local models via Ollama, which is one way to reduce AI vendor lock-in.

The company was founded by Yuhong Sun and Chris Weaver, is based in San Francisco, and has roughly 35 employees. It raised a 10 million US-Dollar seed round in March 2025 co-led by Khosla Ventures and First Round Capital, with Y Combinator participating. The product was called Danswer until the 2025 rebrand, and the legal entity in the repository copyright notice is still DanswerAI, Inc.

One detail worth knowing if you are reading older comparisons: nearly every article about Onyx says it runs on Vespa as the search index. The current deployment manifest ships OpenSearch 3.6.0. The migration is visibly in progress in the codebase, so anything written before 2026 describes a stack that is no longer what you deploy.

Onyx is dual-licensed, not MIT-licensed. The root LICENSE file states it plainly: all content under `ee` directories is covered by the Onyx Enterprise License, everything outside is MIT Expat.

That sounds like a footnote until you look at what is actually inside `backend/ee/`. It holds 274 of 3,125 backend Python files, which is only 8.8 percent of the code. But it is the 8.8 percent that enterprise buyers are paying for:

- `ee/onyx/external_permissions/` contains the permission-sync implementations for Box, Canvas, Confluence, GitHub, Gmail, Google Drive, Jira, Salesforce, SharePoint, Slack and Teams
- `ee/onyx/auth/users.py` handles SSO
- `ee/onyx/access/` holds access control and hierarchy logic
- `ee/server/` covers analytics, billing, SCIM, user groups, licensing, query history and rate limits

Read the sentence that every listicle and even Onyx's own G2 entry repeats: Onyx automatically respects document-level access controls from the source applications. That capability is the enterprise-licensed code, a pattern we found across seven open-source AI tools. The free Community Edition does not ship connector permission syncing.

The enterprise licence text is unambiguous about production use. It permits copying and modifying the software for development and testing without a subscription, but production use requires agreement with the Onyx Subscription Terms or a valid Enterprise License for the correct number of user seats.

We are not accusing Onyx of hiding anything. Fast-moving open-source projects drift out of sync with their marketing pages. But if you are building a procurement case, you should know that three of their own sources disagree with each other, and you should get the answer in writing before you sign.

The third row is the one that bites in practice. For a product whose central promise is host it yourself, there is no published minimum RAM, CPU or disk figure anywhere in the quickstart, the README or the deployment manifests. Your IT team is sizing the box by guesswork.

So we measured it. On 4 August 2026 we deployed the standard stack via the documented `docker compose up -d` path on Docker Desktop with its default memory allocation of 3.83 GiB. Pull and start took 24 minutes. Ten of the twelve containers came up healthy. OpenSearch never started at all. It was OOM-killed in a restart loop, exit code 137, ten times before we stopped it.

The reason is visible in the memory figures: the background worker alone claims 1.68 GiB, the API server another 547 MiB, MinIO 238 MiB. The other nine containers consume roughly 2.8 GiB between them, which leaves the search index too little to boot. The component that dies is the one that makes Onyx enterprise search rather than a chat window.

Onyx Cloud Business is 20 US-Dollar per user per month, Enterprise is quote-only, and the Community Edition is free. Against Glean, which multiple vendor analyses put at 45 to 50 US-Dollar per user per month plus a Work AI add-on and a 50,000 to 60,000 US-Dollar minimum annual commitment, that is a genuine order-of-magnitude difference. Glean publishes no pricing at all, so every Glean figure in circulation is second-hand.

The free tier is where the arithmetic gets interesting. Free describes the licence, not the cost. The standard deployment manifest defines twelve services: the API server, a background worker, the web server, two separate model servers for inference and indexing, PostgreSQL, OpenSearch, nginx, certbot, Redis, MinIO and a code interpreter.

That is a small platform team's worth of surface area, and we costed it in detail in what self-hosting really costs. Someone has to patch it, monitor it, upgrade it across breaking changes and restore it when the OpenSearch cluster falls over at month end. Onyx does offer a Lite mode, which the README says runs in under 1 GB of memory by disabling the index, background workers and model servers through Docker Compose profiles. Lite is a chat UI, not enterprise search.

Onyx's documentation states that the platform is SOC 2 Type II and GDPR compliant. The reports are not published: you request them by emailing [email protected]. We found no statement anywhere in the official documentation about EU data residency for Onyx Cloud, and no publicly downloadable data processing agreement.

Here is the part most vendor-written comparisons will not tell you, because they are all selling something. Self-hosted Onyx solves European data sovereignty better than any SaaS contract can, including ours. If the software runs on a server in Frankfurt that you control, with a local model, there is no international transfer to assess, no subprocessor chain to audit and no adequacy decision to worry about. A DPA is a promise about someone else's infrastructure. Self-hosting removes the need for the promise.

That is the strongest argument for Onyx, and it is worth taking seriously. The trade is operational: you exchange contractual risk for operational risk. Whether that is a good trade depends entirely on whether you have a platform team who wants to own twelve services.

These four are not the same product, which is exactly why buyers get confused. Onyx and Glean compete directly on enterprise document search. Open WebUI is a chat interface without the indexing layer. An AI context layer solves a different problem: it does not search your documents, it understands your organisation.

Onyx is a genuinely strong piece of engineering and the most credible open-source alternative to Glean in 2026. The MIT core is substantial and really is self-hostable. Nothing here is a scandal.

But the story the internet tells about it is wrong in one specific, expensive way. The permission layer that makes enterprise search safe to deploy is not part of the free edition, and Onyx's own pricing page implies otherwise. If you are budgeting on the assumption that free and permission-aware describes the same tier, your business case is wrong by the cost of an Enterprise contract.

The deeper question is what you are trying to solve. Onyx, Glean and Open WebUI all answer variations of find me the document. That is a real problem and worth solving. It is not the same as knowing whether your teams are actually functioning, where knowledge is quietly walking out the door, or which manager is about to lose half their department. Document search will never tell you that, no matter how many connectors it has.