Self-hosted AI workspace that refuses to favor one LLM

What it does

Onyx is a self-hosted AI platform that sits between your team and whatever large language model you prefer. It bundles chat, retrieval-augmented generation across 50+ connectors, agentic workflows, web search, sandboxed code execution, and even image generation into one interface. The project splits into a lightweight Lite mode for simple chat and a full Standard deployment with vector indexes, background workers, and caching for serious workloads.

The interesting bit

Rather than just proxying API calls, Onyx builds an entire application layer around LLMs, complete with an in-house web crawler and deep-research pipelines that the authors claim top their own benchmark. Enterprise features like SSO, RBAC, and audit trails are baked in, yet the core remains MIT-licensed—a rarer combination than it should be.

Key highlights

• Supports essentially every LLM provider, from Ollama and vLLM to OpenAI and Anthropic
• Agentic RAG with hybrid search plus custom agents that can call external tools via MCP
• Sandboxed code execution, voice mode, and artifact generation inside the chat interface
• Two deployment personalities: a sub-1 GB Lite chat UI or a full-scale stack with Redis and MinIO
• MIT-licensed community edition covers core features; enterprise tier adds compliance and whitelabeling

Caveats

• The README cites a deep-research benchmark win dated February 2026, which has not happened yet, so treat that claim with appropriate skepticism
• Standard mode requires a nontrivial infrastructure stack—vector indexes, job queues, and blob stores—compared with the single-container Lite variant
• Enterprise features such as SSO, SCIM, and analytics are reserved for the commercial edition, not the MIT-licensed core

Verdict

Organizations that need a model-agnostic, self-hosted AI workspace with real RAG and governance should evaluate Onyx; individual hackers looking for a minimal chat frontend will likely find it too broad.