The view in the chat

The view sits on a plain background that follows the host's light or dark theme, so it reads as part of the conversation. In Claude Code, the plugin's /things pane plays a similar part (see Related surfaces).

Connect first: Claude or ChatGPT & Codex.

get_workspace_view draws a workspace's top-level things.

• The workspace's name sits at the top. Click it to pick another workspace; with only one, it is a plain label.
• Each thing is a floating emoji tile with its name under it. A thing with several emojis shows up to three, layered; a thing with none shows 📄.
• Names longer than 30 characters are cut short. Hover a tile to read the whole name.
• When the workspace holds more things than the answer carried, a line under the tiles says how many are showing, such as Showing 100 of 240.
• Open this workspace in the app ↗ sits below the tiles.
• An empty workspace reads No things here yet.

The view opens on the workspace the agent asked for, else your default workspace, else your first one.

get_thing_view draws one thing's page.

A question an agent asked about the thing shows between the content and the sub-things (see Questions in the view).

• A tile opens that thing in place, with a new get_thing_view call.
• A breadcrumb step walks back up: the workspace returns to its tiles, and a parent thing opens its own page. The breadcrumb is the way back; the view has no close button.
• The workspace name on the tile screen switches to another workspace, with a new get_workspace_view call.
• open in app ↗ opens the thing in Everythings. The view hands every link to the host to open, the links in a thing's content included.

While a click loads, the screen dims a little. If a read fails, a notice reads Couldn't load that with the reason, and the screen stays as it was. When the agent opened a thing directly, the view makes one small extra read of your workspaces so the picker is ready on the way back.

All of these calls only read, so browsing works on a read-only connection.

• The box flips at once, and the view saves the thing's whole content through update_thing.
• The save goes through the agent's connection, so it needs read & write access, and the app records the tick as that agent's edit.
• Ticks made in quick succession go out in order, and each save carries the latest state on screen.
• When a save is refused (a read-only connection, a host that declined the call, a thing that is gone), the box goes back and a notice reads Couldn't save that with the reason.

• An open question reads “✦ agent is waiting on you”, then the question and its options as a list, then a line pointing you to the app, your phone or the chat.
• Earlier rounds of the same thread sit above it, with who said what and when.
• Once answered, it says who chose which option and where (on the web, on the phone, or in a chat), with any note they added. A dismissed question says so, and a thread the agent closed reads Resolved by the agent.

The view has no buttons for a question. An answer from it would go through the agent's own connection and read as the agent answering itself. Answer on the thing in the app, on your phone, or tell the agent in the chat, and it records your words with answer_request. More in Questions, inbox & @Agent.

• When you ask to see, show or open something: “show me my workspaces”, “open my Recipes workspace”, “show me my music thing”.
• Right after a create or edit you asked for, so the finished thing appears in place for you to confirm or correct.

For its own reading, an agent uses get_thing, list_things, list_child_things and the other plain tools, so a page opens only when you asked for one. When the view is on screen, the agent's reply can skip repeating what it shows. The Tools reference has the full list.