onyx-dot-app/onyx

An open-source, self-hostable AI platform that combines chat, retrieval-augmented generation over your own data, and agents that work with any LLM.

What it is

Onyx is an application layer over LLMs that provides a feature-rich chat interface plus enterprise search across connected data sources. It targets individuals and teams who want a self-hosted assistant with retrieval-augmented generation, custom agents, and actions grounded in their own documents. It works with both self-hosted models (Ollama, LiteLLM, vLLM) and proprietary providers (Anthropic, OpenAI, Gemini).

Key features

Agentic RAG built on a hybrid vector-plus-keyword index for search and answer generation.
A multi-step deep-research flow that produces in-depth reports.
Custom agents with their own instructions, knowledge, and actions, plus actions and MCP for interacting with external applications.
Over 50 indexing connectors out of the box, or connection via MCP.
Web search (Serper, Google PSE, Brave, SearXNG, and others), an in-house crawler with Firecrawl/Exa support, sandboxed code execution, artifact generation, image generation, and voice mode.
Two deployment options — Lite (a lightweight chat UI under 1GB memory) and Standard (full RAG stack with background workers, model inference servers, Redis, and MinIO).

Tech stack

Python backend requiring Python 3.13, built on FastAPI 0.133.1, Pydantic 2.11.7, SQLAlchemy 2.0.50, and Alembic migrations.
Celery 5.5.1 with Redis for background job queues and workers.
LLM access via litellm 1.93.0 and the openai, google-genai, cohere, and voyageai clients.
OpenSearch (opensearch-py 3.2.0) for indexing, with sentence-transformers, torch, and transformers running in a separate model server.
A TypeScript/Next.js frontend and desktop/widget workspaces managed with Bun 1.3.13.
Deployable via Docker, Kubernetes, and Helm/Terraform; enterprise features include SAML/OIDC SSO, SCIM, and RBAC.

When to reach for it

Standing up a self-hosted AI chat and search assistant over internal company knowledge from many sources.
Teams that need enterprise controls such as SSO, RBAC, analytics, and query-history auditing on their AI usage.
Use cases that need RAG plus agents, web search, and code execution in one platform rather than assembled separately.
Staying model-agnostic across self-hosted and proprietary LLM providers.

When not to reach for it

A minimal single-user chat client, where the Standard stack's workers, index, and inference servers are more than needed (Lite mode partly addresses this).
Environments that cannot run the required Python 3.13, OpenSearch, Redis, and model-server components.
Teams wanting a fully hosted service with no operational burden, unless they use Onyx Cloud.
Purely offline single-document Q&A that does not warrant connectors and indexing infrastructure.

Maturity signal

Onyx has over 31,000 stars, dates back to 2023, and is pushed to within the last day, indicating an actively developed and adopted platform. The dual Community/Enterprise editions, extensive connector list, and large Alembic migration history point to a substantial, production-oriented codebase rather than a prototype. The core is MIT-licensed, with enterprise features layered on top.

Alternatives

LibreChat — use instead when you mainly want a multi-provider chat UI without built-in enterprise search and connectors.
Open WebUI — use instead for a lighter self-hosted chat front end, especially over local models, without a full RAG connector stack.
Dify — use instead when you want a low-code builder for LLM apps and workflows rather than a ready-made chat-and-search product.
RAGFlow — use instead when deep document-parsing RAG question answering is your only requirement.

Notes

The repository ships two editions under one license field: an MIT-licensed Community Edition covering chat, RAG, agents, and actions, plus an Enterprise Edition with extra features, which is why GitHub reports the license as unrecognized. The pyproject.toml is unusually heavily commented, pinning specific dependency versions to work around named upstream bugs — for example capping transformers to avoid all-NaN embeddings caused by a from_pretrained() buffer-corruption issue.

Wiki: onyx-dot-app/onyx