The Captra mental model

Captra uses a small set of terms consistently.

Capability

A typed action the application can perform. A capability may be deterministic or backed by an agent, but it always has validated input and output and executes through the same runtime.

Agent

A capability whose implementation is a bounded model/tool loop. An agent adds instructions, a model, tools, and step/token budgets without introducing a second application architecture.

Tool

A reusable typed integration function. Tools declare schemas and permissions and are explicitly registered. Agents can choose tools; they cannot invent or silently replace them.

Hook

Lifecycle middleware around memory, tools, approvals, events, and errors. Transform and policy hooks fail closed. Observer hooks are failure-isolated.

Plugin

A versioned bundle of tools and hooks. Plugins provide the extension boundary for databases, SaaS APIs, policy, and observability.

App

The registry and deployment boundary containing capabilities, tools, plugins, scopes, hooks, and UI surface metadata.

Runtime

The execution engine. It validates contracts, authorizes permissions, recalls memory, coordinates tools, pauses for approval, emits events, persists runs, and handles cancellation and recovery.

View

A transport-safe component name resolved against an explicit local UI registry. The server chooses an allowed view identifier; the client owns the actual component implementation.