The node types on the canvas, the whiteboard and comments.
The canvas is an infinite plane: nodes hold agents and material, links express shared context and collaboration, and groups organize separate tasks.
Node types
Node
Purpose
Terminal (including agents)
A plain shell, or an agent node running Claude Code, Codex or another CLI
Note
Task descriptions, decisions, to-dos; agents can read them along links
Group / Frame
Frames a set of nodes; can be bound to a worktree
Editor
CodeMirror 6 editor with language services and Git gutter marks
Diff
Git diffs of the working tree or between commits
File tree
Browse project files
Browser
One browser session shared by you and your agents
Automation
Cron, interval or one-off schedules
Agent activity
Summarizes an agent’s subtasks and progress
Whiteboard
The whiteboard layer shares the camera and undo stack with nodes: freehand, shapes such as rectangles and ellipses, lines and arrows, text and images. Mermaid flowcharts import as editable whiteboard objects; other diagram types import as images. Whiteboard objects and frames can be “referenced to an agent” and become its material as an image or text.
Comments
Pin comments on the canvas, reply in threads and @-mention people. When an agent reads a node along a link, it also sees the node’s unresolved comments.
Saving and multi-user editing
The canvas saves automatically. On the server shell several people can edit the same canvas in real time, seeing each other’s cursors and selections; offline edits catch up on reconnect.
Status
Agent node status is normalized to four states: working, waiting, blocked (needs you) and done. Data that can’t be obtained is shown as unknown, never as zero or success.
