Onyx is an open source AI application layer that gives teams a self-hostable chat interface with retrieval-augmented generation over their own company data and more than 50 indexing connectors.

What it is

Onyx is the application layer for large language models, distributed as a Python and Next.js project that provides a feature-rich chat interface which anyone can host. It lives in the LLM chat, retrieval-augmented generation and enterprise search ecosystem, and it sits between the models an organisation already runs or rents and the data that organisation already stores. It supports self-hosted providers such as Ollama, LiteLLM and vLLM alongside proprietary providers such as Anthropic, OpenAI and Gemini, which means the model choice stays separate from the interface and the indexing pipeline.

The concrete problem it solves is the assembly work of a production retrieval stack. Most teams otherwise wire together a chat UI, a connector sync pipeline, a vector store, a keyword index and a job runner by hand, then maintain each piece. Onyx replaces that ad hoc combination with one deployable system: a hybrid vector and keyword index for retrieval, background containers that run job queues and workers for syncing knowledge from connectors, and AI model inference servers used during indexing and inference. Advanced capabilities such as RAG, web search, code execution, file creation and deep research are exposed through the same interface rather than bolted on.

Key capabilities

• Agentic RAG that combines a hybrid index with AI agents for information retrieval.
• Deep Research producing in-depth reports through a multi-step research flow, benchmarked on onyx_deep_research_bench.
• More than 50 indexing-based connectors out of the box, plus MCP for connecting further applications.
• Web Search through Serper, Google PSE, Brave and SearXNG, with an in-house web crawler and support for Firecrawl and Exa.
• Custom Agents built with unique instructions, knowledge and actions, where Actions and MCP let agents interact with external applications with flexible authentication options.
• Code Execution inside a sandbox for data analysis, graph rendering and file modification, alongside Artifacts for generating downloadable documents and graphics.
• Voice Mode using text-to-speech and speech-to-text, and Image Generation from user prompts.

Who uses it and how

• Individual users and small teams evaluating quickly run Onyx Lite, which needs under 1GB of memory and a less complex stack, and covers the Chat UI and Agents functionality.
• Larger teams deploy Standard Onyx, which adds the vector and keyword index, background containers for job queues and workers, inference servers, a Redis in-memory cache and a MinIO blob store for performance at scale.
• Enterprises rely on SSO through Google OAuth, OIDC or SAML, group syncing and user provisioning via SCIM, RBAC over agents and actions, analytics broken down by team, LLM or agent, and query history for auditing adoption.
• Security and compliance groups use custom code to remove PII, reject sensitive queries or run custom analysis, and administrators can whitelabel naming, icons and banners.
• Platform teams deploy through Docker, Kubernetes, Helm and Terraform, with guides for major cloud providers.

Getting started

Deploy with a single command: curl -fsSL https://onyx.app/install_onyx.sh | bash. A hosted option, Onyx Cloud, is available at cloud.onyx.app/signup for trying the platform without deploying.

How it compares

The provided facts name no paid products that Onyx replaces, so no licence, hosting or cost-model comparison can be drawn. In this registry, Onyx stands alone as the entry covering unified AI-powered enterprise search.

When to use it — and when not to

A self-hoster running Standard Onyx must operate the vector and keyword index, the background job queues and connector workers, the model inference servers, Redis and MinIO; Lite mode removes those components but also removes RAG and connector syncing. Teams that only need a simple chat front end, or that cannot run containers and object storage, should look elsewhere. The licence also deserves attention before adoption: the GitHub metadata reports NOASSERTION while the README describes a Community Edition under the MIT licence covering Chat, RAG, Agents and Actions, with a second edition whose description is cut off — verify the terms directly. The repository also carries 451 open issues, which signals active but unsettled development.