Core concepts

A small map first; every feature page comes back to these words.

One diagram

Session
  └─ Turn (you send a message → agent runs → reply)
       ├─ Agent: role, prompts, tools
       ├─ LLM: which model handles this turn
       ├─ Skill: workflow and scripts, loaded on demand
       ├─ MCP: external tools and data sources
       └─ Sandbox: where commands actually run

Session

One independent working context. Messages, attachments and execution records belong to a session. Sessions do not affect each other, so you can use different agents and models in each.

Turn

From your message to the agent’s reply is one turn. A turn may contain several steps: tool calls, commands, file edits, or a question back to you.

Agent

A reusable role configuration: system prompts, bound model, built-in tools, MCP servers, skills and optional user prompts. Picking an agent when creating a session brings all of that in.

Skill

A packaged way of doing something: instructions, optional scripts, optional reference files. Skills do not sit in the context until needed, which is why you can keep many of them.

MCP

Model Context Protocol, used to connect external tools and data sources. A server exposes tools; configure it once, then select it per agent.

Sandbox

The environment where commands and file operations actually run. Each command runs isolated; you can preview output, inspect logs and discard the sandbox at any time. Without a sandbox an agent can only talk.

Exchange Space

Hand secrets, variables and files to the runtime safely: injected as environment variables or files, never written into prompts or transcripts.

Prompts

Reusable prompt fragments, split into system and user prompts. Compose them in an agent instead of rewriting each time.

How they fit together

You want to…	Use
Change how the agent talks	Agent + prompts
Let it calculate, search or call services	Built-in tools + MCP
Make it follow a fixed workflow	Skills
Let it run commands and edit files	Sandbox
Give it private credentials	Exchange Space
Switch to a cheaper or stronger model	LLM config