Assistants, agents, and models
Three separate things decide how a conversation behaves. Keeping them apart makes everything else in Pavii.Ai easier to reason about.
Agent — the runtime
Section titled “Agent — the runtime”The agent is the engine that actually executes a turn: reading files, editing code, running commands. Pavii.Ai talks to agents over ACP, a protocol for driving CLI coding agents as subprocesses.
- OpenCode ships inside the installer and is the default.
- Pavii CLI is the built-in agent that needs no external tool.
- Other CLI vendors — Claude Code, Codex, Qwen, Goose, CodeBuddy and more are available as agent adapters from the Agent Hub, and detected if you already have them installed.
Assistant — the persona
Section titled “Assistant — the persona”An assistant is what you pick in the UI. It binds together:
- an agent to run on,
- rules and instructions that shape its behaviour,
- a set of enabled skills,
- selected MCP servers,
- and default choices for model, permission mode, and reasoning depth.
Several assistants can share one agent. That is the point: a “reviewer” and an “implementer” can both run on OpenCode while behaving very differently.
Model — the LLM
Section titled “Model — the LLM”The model is chosen per conversation from the header selector, using your own API key or one of the free provider options.
Auto versus fixed defaults
Section titled “Auto versus fixed defaults”Each assistant default — model, permission mode, reasoning depth, skills, MCP servers — is either:
- auto — follow whatever you last used globally, so the assistant adapts to how you are working right now, or
- fixed — pinned to this assistant, so it always starts the same way.
Built-in assistants pin their skills and leave the rest on auto. Pinning the model is useful when an assistant is tuned for a specific one; leaving it auto is better for a general-purpose helper.
How this plays out
Section titled “How this plays out”Because the three layers are independent, two conversations can run different agents on different models at the same time, and a team can mix agents — one member on Claude Code, another on OpenCode, another on Pavii CLI — inside a single shared goal.