AI chat assistant
The editor has a built-in AI assistant that reads your design and edits it for you. It is provider agnostic and runs on your own key, never a hosted key. Ask it in plain language and it can add or change objects on the canvas, generate an image, or simply answer a design question in the chat. See Bring your own AI key.
Quick overview
- Open the assistant from its icon in the left rail, or press
Ctrl + Shift + I. - Connect a provider and a model in AI settings if you have not already.
- Type a request. The assistant replies and, when the request calls for it, edits the canvas directly.
- Attach an image, paste one, or capture a canvas screenshot to give it something to look at.
- Review the actions it took, listed under the reply, and undo them like any other edit if needed.
- Switch between past threads, or start a new one, from the chat history sidebar.
Detailed reference
Providers
Vision (image understanding) works on every model for OpenAI, Google Gemini, Anthropic Claude, and a custom endpoint. Qwen, Kimi, and GLM only support it on specific models, and DeepSeek does not support it at all.
| Provider | Vision | Image generation |
|---|---|---|
| Qwen (DashScope) | On VL models | Native |
| Kimi (Moonshot) | On select models | No |
| DeepSeek | Not supported | No |
| GLM (Zhipu) | On GLM-4V models | No |
| OpenAI | Every model | Native (DALL-E) |
| Google Gemini | Every model | Native (Imagen) |
| Anthropic Claude | Every model | No |
| Custom endpoint | Depends on the endpoint | No |
Only Qwen, OpenAI, and Google generate images natively. Ask for an image on any other provider and the assistant falls back to a saved key for one of those three, if you have one connected. A custom endpoint must accept OpenAI-compatible chat completions requests.
Canvas actions
Ask for a change and the assistant replies with a short explanation plus a batch of actions, which the editor runs immediately against the canvas. Executed actions are listed under a collapsible note beneath the reply, and because they run through the same code path as your own edits, they undo and redo normally. The assistant works in small batches rather than one giant leap, so a large request can take a couple of follow-up messages to finish.
| Category | The assistant can | Example request |
|---|---|---|
| Content | Add text, shapes, icons, and freeform paths | "Add a bold headline that says Grand Opening" |
| Images | Generate an image from a prompt, or search stock photos and place one | "Generate a photo of a coffee cup on a marble counter" |
| Canvas | Set a background color or image, resize the canvas, switch product type, apply a template, or clear it | "Make this an Instagram Story with a dark purple background" |
| Selection | Change, delete, duplicate, or describe the selected object | "Make the selected text bigger and rotate it 15 degrees" |
| Arrangement | Align, reorder, group, ungroup, or select everything | "Group these three shapes and bring them to the front" |
| Patterns | Build a repeating text pattern, a centered watermark, or a grid of duplicates | "Add a diagonal watermark that says DRAFT" |
| Variations | Read a page's Design Fields, roles assigned from the properties panel, and generate several content variations from them | "Make 10 variations of this quote card with different authors" |
| History | Undo or redo, or report on canvas state | "Undo that" |
Vision and attachments
Attach an image with the paperclip button, drag files onto the chat, paste from your clipboard, or capture the current canvas with the screenshot button. All of these accept more than one image at a time. Images are resized before sending and shown as thumbnails on the message, click one to view it full size. If your selected model cannot see images but the provider offers one that can, the editor switches to it for that message and lets you know. If the provider has no vision-capable model at all, you are prompted to pick a provider that does.
Conversations
Every chat lives in a history sidebar opened from the panel header. Start a new thread with the plus button, switch between threads by clicking one, or delete a thread you no longer need. The first exchange in a thread sets its title automatically. Recent messages in the active thread are sent back to the provider as context on every new message.
Message actions
| Action | Effect |
|---|---|
| Copy | Copies the assistant's reply text to your clipboard |
| Apply to canvas | Drops the reply text onto the canvas as a new text box |
Tips
Small requests, fast results
The assistant edits in small batches and summarizes what it did afterward. For a large composition, ask in steps: set up the background and layout first, then add copy and icons next.
Vision switches automatically
Attach an image while a text-only model is active, and the editor swaps in a vision-capable model from the same provider for that message, when one is available.
Provider keys stay on the server
For every built-in provider (OpenAI, Claude, Gemini, Qwen, DeepSeek, Kimi, GLM), your key is encrypted at rest in your account and never returned to the browser: the editor calls the provider through the server, which injects the key. The one exception is a custom endpoint you add yourself, pointing at a URL you control, whose key is stored locally in your browser and called directly. See Bring your own AI key.
Related
- AI skills for personas that change how the assistant behaves
- Properties panel for assigning the Design Fields that variations read from
- Grouping and layers for how grouped and stacked objects behave
- AI Studio for the full chat, image, and video experience outside the editor
- Bring your own AI key for how provider keys are connected and protected