Onyx
Onyx
Open-source AI chat and search over your team's docs and apps
Onyx is an open-source AI chat and enterprise search platform that connects an LLM to a company's documents, applications and people. It combines hybrid search, advanced RAG, contextual retrieval and LLM-based knowledge graphs to ground answers in team knowledge, and ships deep research, MCP actions, a code interpreter, web search, artifacts, voice mode and image generation. It supports self-hosted model runtimes such as Ollama, LiteLLM and vLLM alongside Anthropic, OpenAI and Gemini, and can be deployed with a single install command or used as hosted Onyx Cloud.
Primary Functions
enterprise search
AI chat
retrieval-augmented generation
custom agents
deep research
Features
- Agentic RAG combining a hybrid index with AI agents for information retrieval
- Deep Research producing in-depth reports through a multi-step research flow
- Indexing-based connectors provided out of the box or via MCP, syncing updates in real time
- Custom agents with their own instructions, knowledge and actions
- Code execution in a sandbox, artifact generation, web search, voice mode and image generation
- Works with self-hosted model providers (Ollama, LiteLLM, vLLM) and proprietary ones (Anthropic, OpenAI, Gemini)
- RBAC and permission inheritance, chat sharing, user feedback and usage analytics
- Deployment via Docker, Kubernetes and Helm/Terraform, plus a lightweight 'Onyx Lite' mode that runs under 1GB memory
Limitations
- Enterprise Edition features live under 'ee' directories governed by the separate proprietary Onyx Enterprise License, not the MIT license that covers the Community Edition
- Onyx Cloud has no free tier: the only published plan is Business at $20 per user per month on annual billing, and Enterprise pricing is quote-only
- Onyx Lite drops the vector plus keyword RAG index, the connector sync workers and the model inference servers, so it is essentially a chat UI
- The homepage RAG comparison is the vendor's own 99-question, 220K-document set scored by two LLM judges, and the repository README says the agentic-RAG benchmark is still 'to release soon'
- Connector counts disagree between vendor surfaces: '40+ App Connectors' on the pricing page versus 'over 50+ indexing based connectors' in the repository README