The short answer
Self-host an open-source enterprise search like Onyx (formerly Danswer) when data residency and full control are hard requirements and you have an ops team to run it. Buy a managed tool when you want cited, permission-aware answers working this week without owning servers, scaling, upgrades or LLM keys. The software is often free; the running of it is not. The real decision is who carries the operational and security burden, because the two features regulated buyers care about most, automatic permission sync and single sign-on, are usually the first things gated behind a paid tier in the open-source products.
Last updated: July 2026.
Open-source enterprise search has gotten genuinely good. Onyx, which was called Danswer until late 2024, is MIT-licensed, self-hostable on Docker or Kubernetes, model-agnostic, and has a large, active community with real deployments behind it. If you have ever wanted an internal search that reads across your apps and writes a cited answer, and you did not want to hand your data to a vendor to get it, that option now exists and it is free to download.
So the question is not whether open-source enterprise search works. It does. The question is whether self-hosting it is cheaper and better for your team than buying a managed one, and that answer depends almost entirely on costs that never appear on a license page.
What "free" actually covers, and what it does not
The license is free. The service is not. When you self-host, you inherit a running system, and someone on your team owns it from then on. Here is the honest ledger.
| Cost | Self-hosted open source | Managed SaaS |
|---|---|---|
| Software license | Free (Community, MIT) | Per seat |
| Hosting and scaling | Yours: servers, a vector store, monitoring, uptime | Included |
| LLM inference | Yours: bring and pay for your own API keys or host models | Included |
| Upgrades and patching | Yours: connectors and dependencies break on version bumps | Handled for you |
| Permission sync and SSO | Often a paid Enterprise tier, not the free build | Usually standard on every plan |
That last row is the one most evaluations miss. In Onyx, for example, the automatic permission (ACL) sync that keeps answers scoped to what each person can see, along with SSO and role-based access, is reserved for the paid Enterprise edition rather than the free Community build. So "free and permission-aware" is not quite true out of the box: the free version searches and answers, but the security machinery a company-wide rollout depends on sits behind the commercial tier.
When self-hosting is the right call
Plenty of teams should self-host, and it is not a consolation prize. Choose it when one of these is genuinely true:
- Data cannot leave your infrastructure. Air-gapped environments, sovereignty requirements, or a security posture that forbids sending document text to a third party. Open source running in your own cloud or on-prem is the clean answer here.
- You want to customize the pipeline. If retrieval logic, ranking, or the agent behavior is something you intend to modify and own, having the source is worth a great deal.
- You already have the ops muscle. A platform or infrastructure team that runs Kubernetes services in production can absorb another one without it becoming a second job for an engineer who has a first job.
The common thread is that you are trading money and time for control, and you have somewhere to put that trade. Someone still has to provision and maintain the servers, keep them patched, and roll out every upgrade without breaking the connectors, which is the recurring cost people underestimate when the download is free.
When buying a managed tool wins
Buy when the goal is the answer, not the infrastructure. Most teams evaluating internal search want employees to stop losing an hour hunting across Slack, Drive and Confluence, and they want that to be true next week, not next quarter. A managed tool gets you there because the vendor already solved hosting, scaling, model access and upgrades, and because the security features you actually need are typically on by default.
This is the honest wedge between an open-source project and a hosted product. It is rarely about capability. Onyx covers the same apps and does cited, permission-aware answers too. It is about who runs it and where the security lives. A managed Onyx alternative like InSearch delivers the same cross-app, cited answers as a service, with item-level permission sync and single sign-on included on every plan rather than as a paid add-on, and nothing to operate.
Is open-source enterprise search secure?
It can be, and self-hosting genuinely helps with data residency because nothing leaves your network. But security is not only about where the data sits. It is also about whether search results respect who is allowed to see what, on every query, for every person, as teams and permissions change. That enforcement is the hard part of internal search, harder than the AI, and in the open-source products it is frequently the feature that lives in the paid tier. So "self-hosted" and "secure by default" are not the same claim. Price the Enterprise edition before you conclude the free version is enough.
A simple way to decide
Ask two questions. First: does your data have to stay inside your own infrastructure, as a hard requirement rather than a preference? If yes, lean toward self-hosting open source. Second: do you have a team that will own the servers, the upgrades and the LLM bill as an ongoing service, not a side project? If no, the total cost of a "free" tool is higher than a per-seat subscription, and a managed product will be live and safe far sooner.
Most teams answer no to the first and no to the second, which is why they end up buying. The ones who answer yes to both self-host, and they are right to. The mistake is answering the license-price question and skipping the operating-cost one, because that is where the real number is.
If you want to see what the managed version of this looks like on your own stack, InSearch searches across all your company apps in one query and returns a cited answer scoped to exactly what each person can open, with no servers to run. It is the same outcome the open-source route promises, minus the operations.