Onyx

Onyx, formerly Danswer, is an open-source enterprise search and AI work platform. More than 40 connectors synchronize content from sources such as Confluence, Google Drive, Slack, Jira, Notion, SharePoint, GitHub, Salesforce, Zendesk, and object storage. Onyx creates vector and keyword indexes over that material and exposes it through search, Chat, custom Agents, Deep Research, code execution, and Actions. Unlike a tool designed mainly for uploading several PDFs, its defining work is the connector lifecycle, retrieval quality, and respect for source permissions.

Quick Verdict

Onyx belongs on the shortlist when employees cannot find decisions, policies, support history, or project knowledge scattered across business applications. It is more connector- and permission-oriented than AnythingLLM, and more search-oriented than LibreChat. The difficult part is not launching a Docker stack. It is assigning least-privilege credentials to every connector, validating that access changes and deletions reach the index promptly, and sizing workers, indexes, storage, and model services for the real corpus.

Best For

The product fits medium and large organizations with several knowledge systems, measurable search friction, and an IT or data team prepared to operate retrieval infrastructure. Engineering can search repositories, issues, and technical documents; support can combine manuals with ticket history; sales can retrieve approved collateral and CRM knowledge; universities can unify institutional sources. A small team that only wants private file chat will find AnythingLLM easier. A team building controlled transactional service conversations should evaluate Rasa. Developers who want a retrieval library rather than a finished employee product can compare LlamaIndex.

Key Features

• More than 40 connectors: Sources span wikis, drives, chat, tickets, code, CRM, meeting transcripts, websites, and object stores. Administrators can schedule synchronization and inspect indexing status.
• Enterprise retrieval: Standard mode includes vector and keyword indexing, background workers for connectors, model-inference services used during indexing and retrieval, Redis, and object-storage components.
• Search, Chat, and Agents: Employees can retrieve source material, ask grounded questions, create custom agents, and enable Deep Research, web search, code analysis, and image features.
• Actions: MCP and OpenAPI tools let agents reach external systems. A code interpreter provides a sandboxed Python execution route.
• Identity and permissions: Business includes basic authentication, Google OAuth, RBAC, permission inheritance, and encrypted secrets. Enterprise adds OIDC/SAML SSO and additional organizational deployment controls.
• Model choice: Administrators can configure major cloud providers, compatible endpoints, LiteLLM gateways, and local systems including Ollama and LM Studio.
• Multiple deployment models: The project supports open-source self-hosting, managed Onyx Cloud, and self-hosted Enterprise Edition.

Use Cases

An employee can search Confluence, Slack, and Jira together to reconstruct why a technical decision was made. Support staff can generate cited answers from product documentation and resolved tickets. Engineering teams can investigate repositories alongside specifications and issue history. Deep Research can synthesize several internal sources rather than relying on public web results. Actions can turn knowledge into work, but read permission and write permission should be designed separately. A safe rollout begins with search-only connectors, then enables narrowly scoped actions with approval and auditing.

Pricing

Onyx Lite is a deployment mode, not a pricing tier. It uses under 1 GB of memory and includes chat, tools, file uploads, and Projects, but excludes the vector database, connector background workers, and indexing inference servers. A serious enterprise-search evaluation needs Standard mode and capacity planning based on the official resourcing guidance.

Pros

• A broad connector catalog supports continuous synchronization instead of one-time file upload.
• Enterprise search, Chat, Agents, and Actions share the same internal knowledge foundation.
• Open-source self-hosting, managed cloud, and Enterprise self-hosting cover different operating models.
• Permission inheritance, RBAC, SSO, audits, and enterprise support have an explicit upgrade path.
• Cloud and local model choices let teams balance performance, data policy, and cost.
• The Lite/Standard distinction gives users a lightweight chat trial without misrepresenting it as the complete search stack.

Cons

• Standard includes indexes, workers, inference services, Redis, and object storage, so it is materially more complex than a chat UI.
• Business is seat-priced, while model, connector, and indexing use still need cost analysis at scale.
• Some identity, region, support, and enterprise-governance requirements need the custom Enterprise plan.
• Connector credentials may read a large amount of sensitive data, creating a high-value secret-management boundary.
• Operators must monitor failed synchronization, stale deletions, index health, backups, and release updates.
• Query logs and retrieved context can still reach external LLM providers unless the full inference path is local and verified.

Alternatives

FAQ

Can Onyx Lite synchronize Confluence and Slack?

Lite should not be evaluated as the complete connector search stack. Official documentation says it excludes the vector database, background workers, and model-inference servers used for indexing. Use Standard for connectors and full RAG.

Does the open-source edition include connectors and Agents?

Yes. Onyx describes the self-hosted open-source option as including core features such as Chat, Agents, Actions, Connectors, and Deep Research. Compare the current edition matrix for specific enterprise identity and permission features.

Why does permission inheritance matter?

An enterprise search result must not reveal a source document that the requesting employee cannot access. Permission inheritance carries source access rules into retrieval. Administrators still need to verify connector support, synchronization delay, identity mapping, and failure behavior for each system.

Can Onyx use local models?

Yes. The documentation covers Ollama, LM Studio, and custom OpenAI-compatible providers in addition to cloud models. Embeddings, reranking, and indexing inference also need review if the requirement is a fully local data path.

Does the $20 Business seat include unlimited model use?

The pricing page lists access to major LLMs but does not establish an unconditional unlimited-use promise on the public page. Ask Onyx to define quotas, overages, fair-use rules, and processing regions before estimating total cost from the seat price alone.

Does self-hosting automatically preserve source permissions?

No. Operators must configure connectors that support permission synchronization, map identities correctly, and test grants, revocations, deletions, and employee departures. A stale index can expose material after the source permission has changed.

Bottom Line

Onyx earns consideration because it makes company knowledge searchable, citable, and permission-aware across many systems, not because it is another ChatGPT shell. Pilot open-source Standard with representative connectors and access-control tests. Choose Business when managed operation is the priority, and enter an Enterprise evaluation when SSO, deployment region, on-premise control, licensing, or dedicated support is non-negotiable.