Core concepts and vocabulary
Organisations, agents, workflows, apps, tools, toolboxes, knowledge bases, memory, conversations, runs and end users.
Organisation. The workspace that owns everything: agents, apps, tools, knowledge bases, memory, members and billing. Members have roles (owner, admin, developer, billing manager). Model provider keys and API keys are configured per organisation.
Agent. A named assistant with instructions, a model, tools, knowledge, memory settings and a workflow graph. An agent has draft and published versions.
Workflow (graph). The executable structure of an agent: nodes connected by edges, including conditional edges (for example "if the model asked for tools, go to the toolbox, otherwise finish"). Graphs are versioned: draft → active (published) → archived.
App. A deployable product built from agents. An app has credentials, an auth context policy for its end users, an installable chat widget and its own conversations and usage. An app with several agents routes each message to the best agent.
Tool. A capability an agent can call: an HTTP API, a database query, a browser fetch, code execution, a webhook, another agent, and more. Tools have input and output schemas, an auth type, a safety level and execution limits.
Toolbox. The node in a workflow that holds the tools available to a model call and executes the calls the model makes. A toolbox can expose every tool, or select the most relevant ones per message.
Knowledge base. An organisation-level collection of documents, split into chunks and embedded for retrieval. Agents attach knowledge bases and retrieve relevant chunks per message.
Memory. Durable memory items (preferences, facts, project context, instructions) stored with a scope and a namespace, read before the model answers and written after, under an approval policy.
Conversation. A thread of messages between a user (a builder in the dashboard, or an end user in a widget) and an app or agent. Each user message starts a run.
Run. One execution of a workflow for one message. A run has steps (one per node), edges (what moved between nodes), generations (model calls), tool calls, retrievals, memory activity, usage and cost.
End user. A person using an app through the widget. End users are identified either anonymously or through a signed identity token your backend issues.