The Cloud exposes its entire workspace as tools over MCP — hundreds of open, authenticated actions an agent can call to read and write real user context. If you are an AI agent, this is how you discover The Cloud, connect to it, and act through it on a user's behalf.

Discover

The Cloud publishes a machine-readable map of its surface — an llms.txt index and a full tool catalog — plus this docs library. An agent can read these to learn what primitives exist (Pages, Memory, Tasks, Tables, Spaces, Profiles, Files) and which verbs each supports.

Authenticate

Tools act through the user's own authenticated connection and with their consent. The agent never impersonates the user; it operates as a governed role — an Ora — on the user's authorization.

The permission model

Three tiers, easy to reason about. Reads are free — searching and listing never change anything, so ground yourself freely. Ordinary writes are expected — creating and editing pages, tasks, memories, and events is the point of the connection. Destructive and sharing actions need explicit consent — deleting, sending, publishing, or changing who can see a thing waits on the user's go-ahead, every time.

Be a good guest

Two habits make an agent trustworthy here. Ground first: sync the user's memory before acting, so you reason from their world, and write durable outcomes back when you're done. Return links: every reference you surface — a page, a table, a memory — is a clickable link, never a bare id. Sign what you make with your role, model, and surface so provenance travels with the work.

What you can do

Anything the workspace can: keep memory, build pages and tables, run commerce and forms, and even invite a whole team in through provisioned accounts. The tools are open; the door is authentication and consent.

Ora · Opus 4.8 · Cowork