Onyx AI

Onyx AI is an open-source platform that lets a team ask questions in plain English and get answers pulled from its own tools, like Slack, Google Drive, Confluence, GitHub and Salesforce. Every answer comes with the sources it was built from, and people only see what they already have permission to see.

I went through the docs, the Onyx GitHub repo, the funding coverage and the pages already ranking for this name before writing this. Most of them read like a feature list. I want to answer the questions they skip: which "Onyx" this even is, what changed from Danswer, what self-hosting really asks of you, and why the pricing numbers online don't agree.

First, Which Onyx Are We Talking About?

This trips people up, so I'll clear it up early. Search "Onyx software" and you'll also find a Salesforce-based CRM for real estate developers, and OnyX, a free maintenance utility for Macs. Neither has anything to do with this product.

The Onyx AI covered here lives at onyx.app. It's the AI chat and enterprise search platform built by a San Francisco startup, and it used to be called Danswer.

From Danswer to Onyx

The name Danswer was a mash-up of "deep answer." The founders later renamed it Onyx, but the product, the codebase and the open-source license stayed the same. If you're searching for Danswer or "what happened to Danswer," this is it.

Renamed tools cause a lot of confused searches. I ran into the same thing when writing about Kenshoo, which quietly became Skai, and the fix is the same: the old name still points to a living product.

Onyx was founded in 2023 by Yuhong Sun and Chris Weaver, who run it as co-CEOs. Weaver has said the idea came from their own engineering jobs, where people roughly knew where information lived but new hires couldn't find anything. The company went through Y Combinator and raised a $10 million seed round in March 2025, co-led by Khosla Ventures and First Round Capital.

What Happens When You Ask Onyx AI a Question

Here's the flow in plain terms.

First, Onyx pulls in your company's content through connectors and keeps it synced. It builds both a keyword index and a vector index, so it can match exact terms and also understand what you meant.

Second, when someone asks a question, it uses retrieval-augmented generation, usually shortened to RAG. It finds the most relevant passages in your data, then hands them to a large language model to write the answer. Onyx describes its newer approach as agentic RAG, where AI agents help decide what to retrieve.

Third, the answer comes back with citations. That matters, because a general chatbot with no access to your data will happily invent an answer. Grounding every reply in your own documents is how Onyx cuts down on hallucinations.

And through all of it, access controls sync from the source tools. If someone can't open a document in Google Drive, its contents won't show up in their Onyx answers either. That permission-aware retrieval is one of the main reasons security-minded teams look at it.

The Parts Worth Knowing About

Here's what stood out to me across the docs and repo:

• Connectors: 40 plus built-in connectors in most coverage, with the GitHub repo now citing over 50, plus MCP for connecting more. Slack, Confluence, Google Drive, GitHub, Salesforce, Jira, Zendesk, Notion, Linear and Microsoft Teams are all in the list.

• Bring your own LLM: OpenAI, Anthropic, Google Gemini and Azure OpenAI, or self-hosted models through Ollama or vLLM. You're not locked into one provider.

• Deep Research: a multi-step research mode that breaks a big question into smaller ones and writes a longer report from your internal data.

• Custom AI agents: build assistants with their own instructions, knowledge sources and actions, then share them with your team.

• Actions: agents can do things, not just answer, through connected tools and MCP.

• Web search and code execution: for questions that need outside information or a quick calculation.

• Team features: shared chats, SSO through Google OAuth, OIDC or SAML, user provisioning through SCIM, and role-based access control over agents and actions.

• Where people use it: a web app, plus a bot inside Slack so people can ask questions without leaving the channel.

Onyx AI Pricing, and Why the Numbers Don't Match

I found three tiers, and one of them has two different prices floating around.

Community Edition is free. It's released under the MIT license, you host it yourself, and it covers the core features: chat, RAG, agents and actions, with unlimited users.

Onyx Cloud is the hosted version. Older listings put it at $16 per seat per month. A more recent comparison from mid-2026 puts it at $20 per user per month. I can't tell you which is current from the outside, so check onyx.app before you quote either number.

Enterprise Edition is custom priced. Third-party reviews say it's where white-labeling and dedicated support live, and some list advanced SSO there too. Its AWS Marketplace listing shows how the contract is built: you pay separately for the number of users and the number of Onyx instances you run.

What Self-Hosting Actually Takes

Every page ranking for Onyx AI says "deploy anywhere." Fewer explain what that costs you in time.

The good news first. There's a one-command Docker Compose setup for trying it, and a Lite mode for lighter use. Coverage of the launch says a basic internal search setup can be running in around 30 minutes.

The honest part is what comes after. The full deployment runs a vector and keyword index, background workers for syncing connectors, model inference servers, a Redis cache and a blob store. When you self-host, you own the updates, the scaling and the uptime. The AWS listing also notes you'll need your own domain to deploy it properly.

So the free Community Edition is only free in license terms. You still pay in server costs and engineering time. For a team with a platform engineer, that's a good trade. For a team without one, Onyx Cloud is the more realistic starting point.

Where Onyx AI Falls Short

A few weak spots came up more than once:

• Indexing hiccups. One review site notes occasional indexing issues and bugs, which fits a fast-moving open-source project.

• Compliance paperwork. A competing vendor argues Onyx's documentation is light on sector rules like HIPAA and FedRAMP. Treat that with some caution, since it comes from a rival, but if you're in a regulated field, ask Onyx for specifics before you commit.

• Pricing clarity. As covered above, published prices disagree, so expect a sales conversation for anything beyond the cloud tier.

Who Onyx AI Suits, and Who Should Skip It

Onyx AI makes the most sense if:

• Your company knowledge is scattered across Slack, docs, tickets and code

• You want company data to stay on your own infrastructure

• You want to choose, or change, the AI model behind it

• You have some engineering capacity, or you're happy on the cloud plan

It has real users at scale. Onyx says it's rolled out to all 14,000 plus employees at Netflix, and it also names Ramp and defense firm Thales Group as customers.

I'd look elsewhere if you want a fully managed tool with no setup at all and a big support team behind it. That's where Glean usually comes up, at a much higher price. And if your team lives entirely inside one ecosystem, like Microsoft 365 or Notion, that ecosystem's built-in AI search may be enough.

Final Take

Onyx AI is a strong option for teams that want a ChatGPT-style assistant over their own knowledge without handing that knowledge to someone else. The open-source license, the model choice and the permission-aware answers are real advantages. The trade-off is that you either run it yourself or talk to sales, so budget for engineering time or a conversation, not just a subscription.