SoAI brings local inference engines and remote OpenAI-compatible providers into one model pool. You manage that pool from a browser, and applications connect through OpenAI-compatible or Anthropic-compatible APIs instead of keeping a separate configuration for each backend.
It is also a workspace for using those models. The WebUI includes chat, files, document collections, agent tools, and scheduled automations, with shared administration for users, access, hardware, and system health. You can run SoAI for one person on a workstation or make it available to a team on a private network.
SoAI is source available, free for personal use, and its complete source code can be inspected in the GitHub repository.
Documentation version policy
This public documentation describes the latest SoAI release. It does not retain separate version-specific documentation sets.
Models and request routing
A plugin connects SoAI to an inference engine or provider. The bundled plugins cover Ollama, vLLM, llama.cpp, translation, embeddings, speech, and any service that exposes an OpenAI-compatible API. Local and remote models then appear together in the model catalog. See the Bundled plugins reference for help choosing a backend.
A new installation has no models. SoAI setup in the first-run wizard can download a local model or connect a remote provider. You can also install a local backend and download a model, register files already on the machine, or connect a provider yourself. The catalog keeps each model's source and capabilities visible, since a chat model, an embedding model, and a speech model do different jobs. The Models catalog page covers each route.
How requests reach a model
Clients send requests to the /v1 API and name a catalog model. The server checks which plugin can handle
the operation, queues the work, starts a stopped local backend when needed, and sends the request when capacity is
available. Clients can keep the same base URL while you change engines or providers behind it.
You can also put several targets behind one virtual model name. It can spread requests across them or use an ordered failover route. Health checks, circuit breakers, and bounded recovery keep a failed plugin from disrupting unrelated backends. Read Resilience and self-healing for the failure states and recovery model.
WebUI, APIs, and agent work
| Surface | Use it for |
|---|---|
| Compatible client APIs | Connect OpenAI-compatible applications and Claude Code to routed models and supported media operations |
| Model Context Protocol | Give MCP clients access to built-in tools, or let assistants use tools from other MCP servers |
| WebUI | Chat, model and plugin management, files, monitoring, administration, and day-to-day work |
Support for an operation depends on the selected model, plugin, and provider. The server presents one API surface, but it does not claim that every backend accepts the same features or request fields.
From chat to agent work
Chat keeps the transcript and its model, prompt, tool, file, and Knowledge settings together. Knowledge collections index your documents so an assistant can retrieve relevant passages during a conversation. See Knowledge and document collections (RAG) for supported workflows.
Built-in tools let an assistant work with the web, files, a browser, shell sessions, mail, calendars, media, and system controls. You choose which tools each conversation mode receives and whether protected actions need approval. The Agent tools overview explains the controls before listing the tool families.
Skills give ordinary chats reusable task instructions and supporting material. Import a skill into your account library, select it for a conversation, and save. The assistant can load it when relevant, or you can request it by name. Creating a skill explains how to package your own procedures and templates.
Products and setup
SoAI Core is the application edition for Linux, Windows, and macOS (experimental) use. SoAI OS contains the complete Core product on a Debian appliance and adds host controls for storage, networking, drivers, updates, and operating-system users. The artifact selects the edition; a setting cannot switch Core into the appliance edition. Compare the editions in the edition comparison, then read the appliance overview if you want the WebUI to manage the host as well as AI services.
SoAI products and editions also explains the separate SoAI Connect client, the built-in SoAI Bench capability, and the distribution boundary between Core and the private appliance edition.
Start here
Start with Install SoAI Core, then follow the Quickstart to create the first administrator, add a local model or remote provider, and send a test message. Before you connect another device, review Security hardening and the Hardening checklist. Use the FAQ for common setup questions and the Glossary for unfamiliar product terms.