Design agent
The DS Manager design agent edits your AI design system from a chat. It changes tokens, fonts and component styles on the canvas, with one undo per turn.
The Agent tab is a chat that edits the design system you have open. It works through the same controls you use, so each change appears on the canvas right away and a whole turn undoes in one step.

Start a chat
- Sign in. The agent edits the system synced to your account and saves your chats there.
- Open the Agent tab in the right rail, or press
2. - Type a request and press Enter, or click one of the starter prompts.
An empty chat offers these starters:
- Make it feel warmer
- Use a rounded, friendly style
- Tighten spacing for a denser UI
- Give it a bold, high-contrast look
- Create a custom component: color picker (Pro only)
While the agent works, the thread shows each step as a line with its time, such as "Looked up rules" or "Edited rules", and the note "Don't close this browser tab". The agent runs in your tab, so closing it stops the turn. Press Stop to end a turn early; edits made up to that point stay.
The chat is saved after every step. If you reload or close the tab mid-turn, the thread keeps what the agent did up to its last finished step, with the note "The agent stopped here: the page was reloaded or closed before it finished." While the tab is in the background, the agent can't take a picture of the canvas, and it says so in its reply.
Click a step to open it and see what it did: a token diff with color swatches, light first and then dark, a diff of the rules, the picture of the canvas the agent took, or the step's result. A long result shows its first six lines, and Show all opens the rest. A step that failed says "did not apply".
When the turn ends, its steps fold into one line above the answer, such as "Worked for 18s · 6 steps". A turn you stopped reads, for example, "Stopped after 5s · 4 steps". Click the line to see the steps again.
When the turn ends, a Changes card sums up what it did. Undo on the card takes back the whole turn, and Redo puts it back. The card's buttons work until you reload the page.
Use the chat picker at the top of the tab to go back to an earlier chat. New chat starts a fresh one and Delete this chat removes the open one. The agent names each chat from your first message.
What the agent can change
The agent edits the open system and nothing outside it. It never writes files or generated CSS; every edit is a change you could also make by hand.
- Color tokens, light and dark.
- Foundations: style, base color, accent and chart presets, radius, spacing, letter spacing, fonts, the type scale and font weights, control heights and focus, icon library and icon style, menu color and accent, shadow, and the system name. Fonts can be Google Fonts or fonts you uploaded to the system.
- Component styles: any part of any component, for one variant, size or state (hover, focus, selected, open, disabled, invalid), in light or dark mode. A request about one component stays on that component. A request about a mood or a brand gets a full pass over tokens, foundations and the components the look touches.
- Variants: add, rename or remove variant and size values, create a new axis such as Avatar
shape, and hide a built-in value or show it again. See Variants and states. - Behaviour defaults: what Select, Popover, Dropdown Menu, Tooltip, Combobox, Dialog and Sheet do when code passes nothing, such as where Select opens. See Behaviour.
- Pro Blocks: one block's own styles, or the shared styles every block uses. On the Free plan only the first block of each category can be edited, and the agent says so ("Hero 2 is part of Pro.").
- Custom components (Pro): build a component shadcn/ui does not ship. See Custom components.
- Lint locks: set up which properties agents may change, through a short interview. See Lint locks.
- Comments: read the comment threads on the system, reply to them and resolve them. See Comments.
After an edit the agent takes a picture of the canvas to check its work, up to four times per message, and fixes what looks broken before it replies.
Ask about DS Manager
The agent also answers questions about the app, such as "If I publish, is it public?", "How do I add a dark logo?" or "How do I sync Figma?". It answers in two to four sentences in the language you write in, names the place in the app, and points you to the right steps. For what it doesn't know, it points you to Docs in the account menu. A request that has nothing to do with DS Manager gets one sentence on what the agent can do here.
Comments
The agent can work through the comments people left on your system. It knows how many threads are open on the page you are on and in the whole system, so you can ask it to "go through the comments". It makes the changes, replies in each thread as "Agent" and resolves it.
On Pro, you can also hand it one thread from the canvas:
- Type @agent in a comment or reply and post it, or click Ask agent in a thread.
- The Agent tab opens with "Address this comment." in the composer and a chip that names the comment.
- Press Enter to send. The agent doesn't start until you do.
The agent then replies in the thread and resolves it. Its edits undo in one step, like any other turn.
Point at something
Selection. ⌘-click an element on the canvas, or click Select an element (the pointer button in the composer) and then click the element. It shows as a chip above your message, and the agent gets that element's styles with your request. The chip's X clears it, and Esc cancels picking.
Files. Click Add a file (the paperclip), or paste. The agent reads:
- images (JPEG, PNG, WebP, GIF)
- HTML, CSS, SVG, JSON, Markdown and plain text
You can attach up to four files per message, each up to 10 MB. A PDF can be at most 2.25 MB. Not every model reads PDFs: when the selected model can't, the composer says so before the file is sent.
Links. Paste a URL in your message and the agent reads that page with its stylesheets. For a design system's site it also reads the color, typography and component pages, then matches them.
A screenshot plus "match this" works well. So does a link to your brand's website.
Limits
- One message runs for at most 30 minutes. At that point the thread says "Stopped after 30 minutes. Your changes so far are saved."
- On included AI usage, an account can run 200 turns or 400k output tokens per day. On your own API key the cap is 500 turns or 2M output tokens. Both reset at midnight UTC.
- If the thread says "The agent is resting until tomorrow.", DS Manager has reached its own daily limit. Try again the next day.
AI usage and the free $1
The line under the composer shows what the open chat has cost so far and how much AI usage you have left. An empty chat also shows your balance, for example "You have $1.00 of free AI edits".
Free plan. You get $1 of AI usage once. It unlocks when your account is at least 10 minutes old and owns a system. Until then the tab says what it is waiting for, either "Create a system to unlock your free $1" or a countdown like "Your free $1 unlocks in 6 minutes", and it opens by itself when the time is up.
Pro. Monthly and yearly Pro include $5 of AI usage each month, about 90 turns. Pro also lets you run the agent on your own API keys. See Plans and limits.
Team. On a team's system, turns draw from the team's pool, $5 for each paid seat every month, and not from your own balance. When the pool runs out, the thread says "This team's AI usage for the month is used up. It refills with next month's grant, or the team's owner can add seats." Viewers, and everyone in a team whose plan isn't paid, have no Agent tab on the team's systems. See Teams.
Pro adds $5 of AI usage a month and lets you run the agent on your own Anthropic, OpenAI, xAI or Google API key.
Use your own API keys
On Pro you can run the agent on your own provider account. A Free account sees "Get Pro to access API keys" in this section.
- Open Settings: Settings… in the account menu, ⌘K, or ⌘,.
- Go to API keys. There is one row each for Anthropic, OpenAI, xAI and Google.
- Paste your key and click Add key. DS Manager checks it with the provider before it stores it. Keys are stored encrypted and never shown again; the row then reads "Key ending in …" with the last four characters.
- In the Agent tab, open the model menu in the composer's corner. The provider's models now appear marked "your key".
Turns on your key are billed by the provider and never touch your included usage. The models on offer are Claude Opus 5, Claude Fable 5.1, Claude Sonnet 5, GPT-5.6 Sol, GPT-5.6 Terra, Grok 4.6 and Gemini 3.8 Flash. Without a key of your own, the agent runs on Grok 4.6 from your included usage.
Switch a key off to keep it without using it. Replace… swaps in a new key and Remove deletes it. If a provider refuses your key during a turn, DS Manager switches the key off and shows the reason next to its row.
When AI usage runs out
A turn that runs out of usage stops, and the thread says "You are out of AI usage." DS Manager never charges you for more.
- Free: the tab shows Out of AI usage with See plans.
- Pro, monthly or yearly: "You have used this month's AI usage. Pro adds $5 more next month. To keep going now, add your own API key." Add your API key opens Settings. Models on your own key keep working.
- Lifetime accounts: "You have used the AI usage that came with Lifetime. To keep going, add your own API key."