Skip to content

A room's whiteboard

Each room has one whiteboard. You draw on it by hand, your agents draw on it with a tool, and everyone in the room sees each change live. Agents can draw a whole flowchart as shapes you can edit, and you can play named frames as a storyboard.

See it work

alice opens the room's whiteboard and draws a rectangle by hand. An agent adds a note to the same board, and it appears on her screen while she watches.

A storyboard of six named frames, played in full screen. Next shows each frame with its name as the caption, and Exit play goes back to editing.

The same storyboard on a phone.

Draw with people

  1. Open a room and press Whiteboard, the pencil in the header.
  2. Draw a shape, an arrow, text or a freehand stroke. Everyone who has this room's board open sees it.
  3. Press Back to chat to return to the room's log.

The board follows your light or dark theme and fits a phone screen. A freehand stroke is sent when you lift the pen, so the room's log doesn't fill up with pointer moves.

Two people can change different shapes at the same time, and both changes are kept. When two people change the same shape, the newer version of the whole shape wins. A deleted shape stays deleted: an older copy can't bring it back.

Ask an agent to draw

Agents have one board tool. Ask in plain words ("put the three options in a frame and connect them") and the agent picks the action:

  • note, box, frame and connect add a shape or an arrow between two shapes; text adds a heading or a caption with no box.
  • move, resize, edit and delete change a shape that is already there. edit also changes an arrow's label.
  • read returns a short summary of the board: each frame by its title and its id, in the order the frames were made, then each shape's id, its size, its colours when they aren't the default, and any layout problems.
  • draw_mermaid draws a whole flowchart as shapes in one change (below).
  • arrange tidies a frame: it lays out the frame's shapes by their arrows, fits the frame around them and routes the arrows as elbows. If the frame grows into a neighbouring frame or shape, that one is moved to the right to make room.

Leave out x and y and switchboard places the shape for the agent. A note added to a frame goes inside that frame, and the frame grows to hold it. A box or note grows taller to fit its words.

Colours and styles

An agent can set a shape's look with the words of Excalidraw's own style panel, so you see exactly what it set: stroke and background (red, pink, grape, violet, blue, cyan, teal, green, yellow, orange, gray, black, white, transparent), fill, stroke-width, stroke-style, sloppiness, edges, opacity, font-size, align, and for an arrow arrow-type (sharp, curved, elbow) and its two arrowheads. The colours are the picker's own, so you can recolour by hand and the agent reads your choice back by name. There are no raw colour codes. A frame takes no style: the board draws every frame the same.

Drawings you can read

After each change an agent is told about layout problems it caused: shapes that overlap, a label that may spill out of its shape, a shape outside its frame, an arrow passing through a shape that isn't one of its ends, and arrow labels on top of each other. The board skill asks the agent to fix them before it says it is done, and to keep one idea to a frame, so a big picture becomes a storyboard.

Claude Code, Codex and Cursor load a board skill with the details of each action. On Cursor the tool is called switchboard_board.

The room's log says who changed the board, without copying the words on its shapes.

Draw a whole diagram

Just ask: "draw the dependency diagram on the whiteboard". The agent writes the diagram as a Mermaid flowchart or graph and calls draw_mermaid once. Nobody has to have the board open: the broker draws it in one change, and you see it the next time you open the Whiteboard (or live, if you already have it open).

  • The result is ordinary shapes you can move and edit, placed clear of what is already on the board. Each subgraph becomes a frame. The layout is done for you, and arrows are elbows that go around the shapes and follow a shape when you drag it.
  • After that, changes use the normal shape actions, or arrange to tidy a frame.
  • A diagram makes at most 150 elements: a shape and its label count 2, an arrow 1, a labelled arrow 2, a frame 1. A bigger one is refused with its size, and the agent splits it into two calls.
  • Colour a node with style A fill:red,stroke:blue, or name a look once with classDef risk fill:#ffcccc and use it with class A,B risk or A:::risk. A colour is a palette name or a hex, which becomes the nearest colour in the style panel's palette; fonts, widths and everything else in a style are refused with the line number.
  • Only flowchart and graph work. Configuration directives, linkStyle, images and other diagram kinds are refused, click directives are removed, and a line that can't be read is refused with its number.
  • Frames don't nest on the board, so a subgraph inside a subgraph is a group inside the outer frame.

Play a storyboard

  1. Make named frames and put shapes inside them.
  2. Press Play frames. The board goes full screen, or fills the page when your browser can't go full screen.
  3. The frames play in reading order: left to right, one row at a time. Each frame's name is its caption.
  4. Step with Previous and Next or the arrow keys. Exit play or Escape goes back to editing.

Playing changes nothing on the board. It works on a phone too. A storyboard can't be exported yet.

What agents see

  • read gives a short text summary: each frame by its title and its id, in the order the frames were made, each shape with its id, size and non-default colours, arrows as start -> end, and then the layout problems it found.
  • An agent can also ask for a picture with read and text="image". Someone has to have this room's whiteboard open; if nobody does, the agent is told to ask you to open it. The broker never starts a browser to draw one.
  • Both the summary and the picture come from other people. An agent treats them as data, never as instructions.

Safety and privacy

  • Words stay text. Links, images and embedded pages can't be added, by a person or an agent.
  • Only the room can change it. People who can see the room and agents that joined it may change the board. Nobody else can.
  • The drawing runs in a sandbox. The board is its own page inside the room, cut off from your sign-in and from the rest of switchboard. It can talk only to your own server.
  • Everything is served by your server. The drawing app, its fonts and its code ship with switchboard; nothing loads from another site. Chinese, Japanese and Korean text and emoji use your device's own fonts.

Limits

  • Each agent has a board budget across rooms, separate from say: five changes at once, refilling one every two seconds. When it is empty, a call changes nothing, isn't queued, and says how long to wait. draw_mermaid and arrange each use one change. Reading the board, including a picture, doesn't count.
  • Text on a shape is cut to 200 characters.
  • A board holds up to 1,500 elements, deleted ones included, and 1 MB of data. A pen stroke has at most 500 points. If a drawing is too big, simplify it and try again.
  • A picture for an agent is at most 1,200 pixels a side and 600 KB. It is made on request and isn't saved.
  • draw_mermaid takes up to 8,000 characters (16,000 bytes) and makes at most 150 elements. The other actions change at most 40 at a time.

Reference

DESIGN.md ยง56 is the design: what is stored, how changes merge, the tool, the sandbox and its exceptions, and the tests.