Working with your AI

Connecting your account

The AI Assistant is a tab in the right dock. Open it and you get a chat panel with a composer at the bottom and, if you have not connected anything yet, an amber prompt to add a key.

1. Open AI Assistant settings Click the gear icon in the panel header. It also opens by itself the first time you try to send a message with no key set.
2. Pick a provider The left nav lists Anthropic (Claude), Google Gemini, OpenAI and Claude Code (local), plus a separate Image generation entry.
3. Paste your key Each provider pane takes your own API key, optionally a model id, and for the OpenAI-compatible providers a base URL. Anthropic defaults to claude-opus-4-8, Gemini to gemini-2.0-flash, OpenAI to gpt-4o.
4. Save and use this provider The header subtitle switches to a green Ready once a key is set. Amber means it still wants one.

There is no backend. The key sits in your browser's local storage and the calls go from the page straight to the provider. Rayzia never sees it, and there is no shared key to run out. The corollary is that the bill is yours and rate limits are yours.

If you run Claude Code locally there is a fourth path with no API key at all. The Claude Code (local) pane gives you a bridge port, a Connect button and a SKILL.md you copy across. It talks to a local MCP server on loopback only, with a capability token that is generated per browser and embedded in the copy of SKILL.md you hand over.

What it can actually do

The assistant works through a layer of semantic verbs, forty-one of them, executed against the real engine. Grouped roughly:

• Draw and select: select, useTool, createShape, drawPath, freehand, spray, group, clearCanvas.
• Text: createText, setText, areaText, textOnPath, touchType, listFonts, loadFont.
• Paint: setPaint, addGradient, widthProfile.
• Effects: applyEffect, applyPathEffect, applyFilter, applyBlend, transformEffect, applyEnvelopeWarp.
• Images: generateImage, placeImage, placeImageFromUrl, searchImages.
• Assets: saveAsset, listAssets, applyAsset, generateAsset, validateAsset, plus the folder operations.
• Housekeeping: setLocked, setHidden, batch, runCommand.

It is not limited to that list. It can enumerate the editor's tools and commands and drive them: useTool activates any tool on the rail, and runCommand reaches engine commands directly. It can also read the document, request a render of the canvas and look at it, and enumerate the available effects, which is why "make this look less flat" gets a considered answer rather than a guess.

Everything the AI does is bracketed as history. Each verb it calls is one undo step, labelled with what it did. A batch of many operations collapses to a single undo. Freehand and pen-style draws replay through the real tools, so those may land as more than one step. Either way, Ctrl+Z gets you back.

Prompting it well

The assistant is competent and literal. Most disappointing results come from a prompt that was vague about the thing that mattered.

• Select first, then talk. If something is selected, say "recolour the selection to a warm palette" and it operates on exactly that. An unselected "recolour it" makes it guess.
• Give numbers when you have them. "Three rectangles, 120 by 40, stacked with 16px gaps" is a job. "Some boxes" is an invitation to improvise.
• Name the outcome, not the technique. It can enumerate its own tools. "Make the badge edge look hand-inked" gives it room to pick the right effect; "apply Roughen at 2" is fine too when you know exactly what you want.
• Iterate in small steps. Ask for the layout, look at it, then ask for the colour. One long instruction is harder to undo cleanly than four short ones.
• Ask it to batch. Fifty similar shapes in one batch is one undo step and one round trip. It is also cheaper on your key.

There is a cap of sixteen steps per message. If it hits that you will see a message saying so. That is usually a sign the request should have been two requests.

What it will not do

The AI boundary carries a deny list, and it is not negotiable from the chat. The assistant cannot open, save, export, import, place files, print, close or reload the document, and it cannot call internal commands. Ordinary editing operations, including delete and clear, stay allowed, because those are undoable and the whole design rests on that.

Image handling is deliberately narrow. Generated images are embedded directly. Images fetched from the web are re-encoded before they enter the document, so no remote URL ends up baked into your file.

You still own the review. The assistant is fast and occasionally confidently wrong, exactly like every other model. Watch the canvas, not the chat log.

Generate is a different thing

There is a second tab called Generate, and it is not the assistant. It makes library assets rather than canvas artwork. Describe what you want, pick a Kind from Brush, Swatch, Gradient, Symbol, Filter, Stylize, Path Effect or Marker, and the result goes through a validation gate before landing in Assets under Yours, in the AI Generated folder.

The gate checks schema, size, numeric ranges, sanitisation and then actually applies the asset briefly to see whether it works. A failure gets one automatic retry using the gate's own report. It reuses the chat provider's key, so connecting once covers both surfaces.

Rule of thumb: AI Assistant changes your document, Generate stocks your toolbox.

Common questions