Yuxi
Self-hosted AI agent platform for document Q&A and multi-agent tasks. Runs with Docker, uses OpenAI-compatible model APIs, and has an MIT license.
Yuxi is a self-hosted knowledge agent platform for teams that want to use their documents in both Q&A and multi-step work. It combines document retrieval, knowledge graphs and agent execution in a browser workspace, with permissions for users and departments. It's open source under MIT.
The platform runs on your own server with Docker. Model inference uses a configured API, with support for OpenAI, DeepSeek, Qwen and custom OpenAI-compatible services. Self-hosting the workspace doesn't make those model services local: choosing a cloud provider sends model requests to that provider.
Knowledge bases accept PDF, Word, PowerPoint, Excel and Markdown files, including text, images and tables extracted by document parsers. Agents retrieve relevant passages for answers with traceable sources. Milvus and Neo4j support vector retrieval and graphs of entities and relationships; Yuxi also connects to Dify and Notion knowledge sources. Retrieval tests and Q&A evaluations help teams check which material an answer draws on and measure answer relevance.
For tasks beyond answering questions, LangGraph coordinates agents and specialized sub-agents, with MCP, Skills and tools extending what they can do. The workspace shows task progress and supports human approval for sensitive operations. Sandboxed tasks can produce persistent files that users preview or download, including documents, spreadsheets, images, PDFs and webpages. Team permissions cover knowledge bases, agents, Skills and models.
