Open WebUI is a self-hosted AI platform that runs entirely offline. It connects to local Ollama models and to any OpenAI-compatible API — LM Studio, GroqCloud, Mistral, OpenRouter, vLLM, and others — so one interface covers local and cloud providers at once. Installation is by pip, uv, Docker, or Kubernetes through kubectl, kustomize, or helm, with :ollama and :cuda tagged container images.

Beyond chat, the platform provides a notes workspace with a rich editor and AI rewriting that can be attached to any conversation, channels where teammates and models collaborate on one timeline with threads and reactions, persistent memory that carries facts between chats, calendars that models manage through function calling, and automations that run prompts on a schedule and link each run back to the chat it produced.

Retrieval-augmented generation runs locally against nine vector databases — ChromaDB, PGVector, Qdrant, Milvus, Elasticsearch, OpenSearch, Pinecone, S3Vector, and Oracle 23ai — with content extraction through Tika, Docling, Document Intelligence, Mistral OCR, PaddleOCR-vl, or external loaders, hybrid BM25 and vector search with reranking, and a full-context mode. Web search for retrieval reaches dozens of providers including SearXNG, Google PSE, Brave, Kagi, Tavily, Perplexity, Firecrawl, DuckDuckGo, Bing, Jina, and Exa. Image generation and editing run through OpenAI DALL·E, Gemini, ComfyUI, or AUTOMATIC1111.

Extensibility comes from filters, actions, pipes, tools, and skills, plus external services reached over MCP, MCPO, and OpenAPI tool servers. Any base model can be wrapped with instructions, tools, and knowledge to become a named agent with its own access control, and presets can be imported from the project's community site.

For teams, administrators define roles, groups, and permissions, with LDAP and Active Directory integration, single sign-on through trusted headers or OAuth, and SCIM 2.0 provisioning for identity providers such as Okta, Azure AD, and Google Workspace. Admin dashboards track message volume, token consumption, and cost per user and per model, and a built-in arena with A/B testing and ELO leaderboards compares models. Deployments store data in SQLite with optional encryption or PostgreSQL, keep files locally or on S3, Google Cloud Storage, or Azure Blob Storage, emit OpenTelemetry traces, metrics, and logs, and scale horizontally behind a load balancer on Redis-backed sessions and WebSockets.

Companion projects extend the core: a standalone mobile-first computer and coding agent, self-hosted terminal environments that let the model write and run code inside a chat with per-user isolated containers, a knowledge-base sync service covering more than 45 sources, and a native desktop application with a system-wide chat bar and an optional built-in llama.cpp engine for local inference.

Features

Provider-agnostic models: local Ollama models alongside any OpenAI-compatible API endpoint
Deployment options: pip, uv, Docker, or Kubernetes via kubectl, kustomize, or helm, with :ollama and :cuda images
Local RAG: nine vector databases, multiple extraction engines, hybrid BM25 plus vector search with reranking, and full-context mode
Web search and browsing: dozens of search providers feed results into a conversation, and pages can be pulled in by URL
Plugin system: filters, actions, pipes, tools, and skills, plus MCP, MCPO, and OpenAPI tool servers
Models and agents: wrap a base model with instructions, tools, and knowledge, with per-user and per-group access control
Collaboration surfaces: notes, real-time channels with threads and reactions, and shared calendars managed by models
Automations: scheduled prompts whose runs appear on the calendar and link back to the resulting chat
Voice and image: speech-to-text and text-to-speech through several engines, and image generation with DALL·E, Gemini, ComfyUI, or AUTOMATIC1111