Onyx

What it is

Onyx — formerly Danswer — is enterprise search and AI chat that you can run on your own hardware. It indexes your SaaS through 50-plus connectors, answers questions against that index with the model of your choice, and exposes the same retrieval layer to agents and MCP actions. The GitHub repo carries 31.9k stars and 4.4k forks, which makes it the default answer whenever someone asks for an open-source Glean.

It is the first fork in most enterprise-search evaluations, and the fork is usually framed wrong. Buyers read “MIT-licensed” and hear “the same product as Glean, for free.” Onyx ships in two editions, and the line between them runs straight through the feature that makes enterprise search safe to turn on.

The open-core line, which is the whole evaluation

The Community Edition is MIT: chat, retrieval, agents, actions, connectors, the API. A fully FOSS build lives in the separate onyx-foss repository. That edition is genuinely free and genuinely useful.

The Enterprise Edition is a commercial license, and it holds the list a security review will ask for — OIDC and SAML SSO, permission sync connectors, user groups and RBAC, group-based permissions on connectors and document sets, usage analytics, encrypted secrets, hook extensions, and priority support.

Permission sync is the one that decides the purchase. It mirrors each source system’s access control list and enforces it at retrieval time, so a compensation doc surfaces for the person who could already open it in Drive and stays invisible to everyone else. Without it, your index is flat: anyone who can log in can retrieve anything you indexed. That is a workable posture for engineering docs and a public handbook. It is not a workable posture for an index that includes HR files, board material, or an unredacted deal room.

So the honest framing is three options, not two. Free self-hosted Onyx over content you are comfortable exposing to every logged-in employee. Paid Onyx — cloud or on-premise — with real ACL enforcement. Or Glean. The middle option is still cheaper than Glean and still yours to host; it is not free.

Pricing reality

Vendor pricing page, checked 5 September 2026:

• Community Edition — $0. Self-hosted, MIT, no seat count, no vendor relationship.
• Business — $20 per user per month on annual billing. Onyx Cloud. Chat and search UI, major LLMs, custom agents, MCP and OpenAPI actions, 40-plus connectors, web search, deep research, code interpreter, Slack integration, developer APIs, RBAC and permission inheritance, community support.
• Enterprise — quote-only. Adds OIDC/SAML SSO, on-premise and region-specific deployment, white-labelling, custom integrations, hook extensions, data exports, dedicated support and an SLA.

Note where on-premise sits: paid self-hosting is an Enterprise-tier conversation, so the “we host it ourselves and skip the sales call” plan only holds for the free edition.

Against the category, Glean runs roughly $40-50 per user per month at a 100-seat minimum, plus about $15 per user for the AI layer when it is broken out. Onyx Business at $20 is a little under half of Glean’s base before the add-on. For 300 seats that is about $72K a year versus a Glean contract that lands in the low hundreds of thousands once onboarding and integration work are counted.

Then add the infrastructure you are now operating. Standard mode wants 4 vCPU and 10 GB RAM minimum, 8-plus vCPU and 16-plus GB recommended, and 32 GB of disk plus roughly 2.5x your indexed data — the docs suggest 500 GB for organizations under 5,000 users. Memory tracks corpus size at about 2 GB per GB of source documents for small corpora, easing to about 1 GB per GB above 50 GB. A Lite mode runs in 2 vCPU and 2 GB for evaluation. None of that is expensive in cloud spend; all of it is somebody’s on-call rotation.

Best for

An engineering-adjacent ops or platform team of 100 to 2,000 seats that already runs its own infrastructure and has a specific reason to keep the index in-house — data residency, a defence or regulated customer base, or a Glean quote that came back above what the problem is worth. Onyx is deployed to all 14,000-plus employees at Netflix and inside Thales, per the vendor’s own published customer material, so the ceiling is not the constraint.

The clean tell: if the person who would run this already operates Postgres and a search cluster and would not consider it a new burden, Onyx is the right shape. If that person does not exist, buy the cloud tier or buy Glean.

Not for

• Teams with no platform owner. Self-hosting an index is a service, not an install.
• Anyone who needs day-one ACL enforcement on a free budget. The two conditions are mutually exclusive here. Pick one.
• Deployments where connector coverage is the binding constraint. Glean’s 100-plus catalog is deeper, and a missing connector for your system of record is not something a licence saves you from.
• Small teams under about 50 seats. At that size the operational overhead outweighs the seat savings and a hosted assistant with a connector layer is the better trade.

Versus the alternatives

• Glean — the share leader and the benchmark Onyx is measured against. Pick Glean when retrieval quality across 100-plus systems is the deciding factor, the permissions model is already clean, and nobody wants to own the infrastructure. Pick Onyx when hosting location is a requirement rather than a preference, or when the quote is the objection.
• Dust — the fast-growing AI-native entrant, agent-first rather than index-first. Pick Dust when the goal is building and shipping assistants to teams quickly and search is a means to that; pick Onyx when the index itself is the asset. The Dust vs Glean comparison covers that axis in depth.
• Hebbia — document-heavy analytical work over a defined corpus rather than company-wide retrieval. Different job.
• Open WebUI or LibreChat — if what you actually want is a self-hosted chat front end and you have no connector requirement, these are lighter and the resource floor is far lower. Onyx earns its footprint through the connector and permission layers; skip it if you will not use them.

If none of these fit, the problem may not be search at all — best knowledge bases for teams covers the case where the content is missing rather than unfindable, and RAG explains what the retrieval layer can and cannot do.

Watch-outs

• The free edition has no permission sync, and pilots hide this. A 20-person pilot indexes engineering docs, everyone can see everything, nobody notices. Rollout adds Drive and the HR folder and the flat index becomes a disclosure incident. Guard: decide the edition before the pilot, not after. If the production index will ever contain something not everyone should read, run the pilot on the Enterprise trial — the vendor offers a free self-hosted one on request — so you are testing the thing you will actually deploy.
• Disk is the failure mode nobody plans for. Budget roughly 2.5x your indexed corpus, and note that the underlying OpenSearch cluster applies a read-only block once disk crosses the flood-stage watermark. That surfaces as indexing silently stopping, and stale answers are worse than no answers because nobody can tell. Guard: alert on index-freshness age, not just disk percentage, and re-forecast after the first full crawl — the first crawl is when you learn your real corpus size.
• Open core moves. The company is venture-backed with a $10M seed co-led by Khosla Ventures and First Round Capital, and the CE/EE line is set by the vendor and can be redrawn. This is not a prediction, it is the structural risk of the model. Guard: pin the version you deploy, track the onyx-foss build as your fallback, and write the migration path into the decision doc rather than discovering it during a renewal.
• Two unrelated companies are named Onyx. An Israeli AI security startup shares the name and raised a $113M Series B in 2026; the ML runtime ONNX is a near-homograph. Guard: search vendor records, security questionnaires and news alerts on onyx.app or “Onyx (Danswer)”, or your security review will be reading another company’s funding history and threat model.
• Admin tooling trails the chat surface. The recurring complaint from operators is document-level visibility — tracing which source produced an answer and what actually got indexed. Guard: during evaluation, take ten questions with known-correct answers and verify you can trace each one back to its source document in the admin UI. If you cannot, you have no debugging loop when retrieval quality drifts.

Related: the enterprise AI rollout stack prices the layers that sit around a retrieval tool like this one.