Working with your assistant

Once your assistant is connected, you work with it in your AI app as usual: describe what you want, and it reads the board and makes the changes live in your browser tab. Keep the board open in a connected browser tab while your assistant edits it:

• Point it at a board and manage settings: choose which board it works on, create or manage saved boards, and adjust grid or snapping settings.
• Show what you mean on the board: select shapes, select an area in the Image Editor, frame a part of the board in the view, or ask for specific shapes such as a table, rich text, or an HTML shape with attributes you can tweak.
• Work with files and images: let your assistant read files on the board, or use its AI app's image generation tools to create and edit images directly on the board.
• Follow its changes: open Agent activity to review each agent's tasks and steps, find the shapes it changed, and block its access when needed.

Pointing it at the right board

Your assistant usually finds the board by its name, for example "In my Product launch board, add a title". If it can't, or several boards share a name, give it the board's ID:

• Operation:
  1. Open Use AI with this board. Once an assistant is connected, the button shows the connection status instead, such as MCP server connected or WebMCP connected.
  2. Open Settings and choose Copy Board ID.
  3. Paste the ID into your request.
• Result: your assistant works on that board.
• Tip: with several board tabs connected, name the board you mean in each request.

Managing boards and settings

Your assistant can manage the boards saved in this browser and change board or editor settings:

• Create, open, rename, or duplicate a board: creating or duplicating a board leaves the current board open until you ask your assistant to open the new one.
• Delete a board: permanently removes the board's data from this browser. Managing boards is outside the board's undo history.
• Grid: ask your assistant to show or hide the Grid or change its spacing. The grid is saved with the board.
• Snapping and tooltips: ask it to turn Snap or tooltips on or off. These preferences apply across boards and are outside the board's undo history.

Showing what you mean

Your assistant can read what you are looking at and what you have selected in the open tab. Reading the tab state pauses animation playback:

• Select shapes: select one or more shapes and ask your assistant to work on "the selected shapes".
• Select an area in the Image Editor: in the Image Editor, select part of an image and ask your assistant to work on "the selected area". Your assistant tells a selected image area apart from selected shapes (see Creating and editing images).
• Frame an area in the view: move or zoom the view to the part of the board you want your assistant to focus on before asking about it.
• Ask for specific shapes: name the shapes you want so your assistant uses the right ones instead of relying only on basic geometric shapes and plain text. For example, ask it to create a table, write formatted notes in rich text, or build an HTML shape with custom attributes in HTML State that you can tweak from the Attributes panel.
• Let it take a picture of the board: your assistant can capture an image of the open board or current page to check its results. Use Share live content to include live content in its screenshots (see Sharing live content).

Files on the board

Your assistant can read the files on the open board, within the connection's size limits, and use them for what you ask, such as a spreadsheet to chart. Which formats it understands depends on your AI app.

• What it can read:
  • The files behind file cards, and pictures, videos, PDFs, and sounds.
  • Attachments in table cells and HTML shapes, and pictures inside rich text.
  • The grid image, and the pictures the board makes from these files, such as a video's cover.
• Check what files are on the board: your assistant can list the board's files with their names, types, and sizes before opening any of them. Some of these details may be missing.
• How a file reaches it: a file the board loads from a web address reaches your assistant as that address. Any other file on a local board stays in your browser, so it travels through the connection:
  • Local board MCP server: gives your assistant a temporary link on your computer to read the file.
  • Board WebMCP: has no local server, so the board sends the whole file inside its reply to your assistant. Only small files fit (see Limits).
• Long documents: your AI app limits how much the board can send it in one reply, so your assistant reads long documents (PDFs, e-books, and comic books) a few pages or chapters at a time, through either connection.
• Programs: a file detected as a program reaches your assistant like any other file, marked as a program. Your AI app decides what to do with it.
• Get a file for yourself: ask your assistant to prepare it on the board, for example as a file card, then export or download it from the app.

Creating and editing images

If your AI app supports image generation, your assistant can create and edit images directly on the board:

• Create new images: ask for an image, and your assistant generates it with its AI app's image tools and places it on the board.
• Edit part of an image: in the Image Editor, select an area and ask your assistant to replace it, or to change everything outside it. It reads the selected area, generates the new image, and updates the image on the board. To keep the original untouched, ask it to put the result on a new layer with a mask. Keep the area selected until it finishes.
• Create masks: ask your assistant to create a mask for a layer in the Image Editor, for example to hide a background.

Generated images or image files on your computer reach the board through the local board MCP server, which reads them from the folders it can access. When Codex generates an image in your conversation, your assistant identifies that generated image before adding it to the board.

Following what it does

The Agent activity line below the connection button lets you follow your assistant's requests on a computer.

• Operation: click the activity line to open or close the dropdown.
• Result: the dropdown opens in place, and you can continue using the board while it is open.
• Current activity: after the first request, the line names the agent and shows its current activity. Between requests, it shows the time of the last action.
• Requests needing your answer: a new request opens the dropdown automatically and comes first, under Waiting for you.
• Retained activity: reloading keeps the activity in this tab. Closing the tab or signing out clears it.

Agents and tasks

Open an agent's card in the activity dropdown to see its tasks, newest first.

• Agent name: the AI app's name and a number identify each connection. The colored dot identifies the app.
• Current activity: each card shows one of these states:
  • Reading the board: a request is reading content.
  • Editing the board: a request is changing content.
  • Waiting for you: a request needs your answer.
  • Time of the last action: shown between requests in your local time. Actions from another day also show the date; those from another year include the year.
• Disconnected agents: the dot becomes muted and is marked Disconnected. Their tasks and the time of the last action remain available.
• Browser agents: assistants using Board WebMCP in the same tab share one Browser agent card.
• Tasks: titles use the agent's own description when it supplies one, or name the action it requested. Click a task to open its steps.
• Steps: a tick marks a successful step; a red cross and a reason mark a failed step. Repeated reads of the same content share one line with a count.
• Step details: click a step to see the affected shapes, named by type and text, grouped as Created, Changed, or Removed. A task with one step opens directly to these details.
• Tool call: each step includes these details:
  • Tool: the tool the agent used.
  • Arguments: what the agent asked it to do.
  • Result: the tool's response, or Error if the request failed.
  • Duration: how long the request took.

Selecting shapes from activity

Use the Select button beside a shape's name to find it on the open board.

• Point at the button: highlights the shape.
• Click: selects the shape.
• Hold Shift and click: adds the shape to your selection.
• Double-click: selects the shape and zooms to it.

Filtering and resizing activity

The dropdown's header and bottom edge control how much activity you see.

• Show edits only: the pencil toggle in the header filters the tasks.
  • On: shows tasks with edits and requests waiting for you.
  • Off: includes read-only tasks too.
• Change the height: drag the bottom edge of the dropdown.
• Fit the content: double-click the bottom edge to fit the activity within the available window height.

Blocking an agent

The Block access toggle on an agent's card controls that connection's access across Irakasi tabs in this browser.

• Operation: open the activity dropdown and turn on Block access on the agent's card.
• On:
  • The agent's name dims and the button turns red.
  • Queued and new requests are refused in every Irakasi tab in this browser. The agent receives the reason.
  • A request already running finishes, and existing changes stay.
• Off: turn the toggle off to allow the agent to use the board again.
• Local MCP connections:
  • The block survives a page reload or a temporary connection loss.
  • It ends when you turn it off or the agent ends its connection.
  • Restarting the AI app starts a new, unblocked connection.
• Board WebMCP: the block applies to the browser agent shared by assistants in this tab. Reloading or leaving the page ends that connection and its block.

Sharing live content

The Share live content toggle beside the connection button shares this tab for your assistant's screenshots.

• Operation:
  1. Turn on Share live content.
  2. Choose this Irakasi tab in the browser's sharing prompt and allow sharing.
• Needs: a connected agent and live content in the board:
  • A web page.
  • A streamed video.
  • An HTML shape.
• On: a green dot marks sharing. Screenshots requested by your assistant include the live content.
• Off: stops sharing the tab with your assistant.
• Waiting for you: the screen button is highlighted when an agent asks to see live content. The request also appears at the top of the activity dropdown.

Undoing its changes

Your assistant's changes share the board's undo history with your own edits:

• Failed requests: when a request fails, some of its changes or files it added may remain on the board. Check the board before asking again.
• Ask your assistant to undo or redo: your assistant uses the same undo history, so it can undo your own edits as well as its own.
• Versions: switching or managing Versions cannot be undone, and switching Versions clears the undo history.

Several assistants on one board

You can connect several AI apps to the same board, through the local board MCP server, Board WebMCP, or both:

• One request at a time: a board tab runs one request at a time, so two assistants never change the board at the same moment.
• Shared board state: each assistant sees the others' changes the next time it reads the board.
• Shared undo history: their edits and yours share the board's undo history.
• Agent activity: each connection has its own card and tasks. Assistants using Board WebMCP in the same tab share one browser agent card (see Agents and tasks).

Limits

The connection limits the size of the files your assistant adds, reads, or exports:

• Images it adds: up to 10 MiB each.
• PDFs and other files it adds: up to 50 MiB each, or 25 MiB when your assistant includes the file in its message.
• Videos and sounds it adds: up to 50 MiB each.
• Files it reads and exports it makes: the limit depends on how your assistant connects:
  • Local board MCP server: up to 25 MiB for an image or an export, or 50 MiB for any other file.
  • Board WebMCP: a whole file that travels inside the board's reply must be smaller than 60 KiB, and an SVG export must have fewer than 40,000 characters. The reply's other details take part of that space.
• Larger files with Board WebMCP: the board refuses to send a whole file that doesn't fit, and your assistant can ask for a smaller export instead. Long documents can still be read a few pages or chapters at a time. To get a larger whole file that has no web address, connect through the local board MCP server.
• Temporary file links on your computer: only an assistant that runs on your computer can open the temporary file links created by the local board MCP server. An assistant that works from its provider's servers cannot open links on your computer.
• Stopped requests: if you stop a request in your AI app after it was sent, the change may still finish on the board. Check the board before asking again.