v2.16.0

Features

• Buttons on the page that ask the agent for something now reach it. While the agent listens with comments <root> --wait, such a button sends its request to the agent the way Send to Agent sends a comment, and the page says it was sent to the agent. --wait prints it under requests: its text is what is asked, and label and selection say what it is about (an --agent-hook command gets {"request": …}). In claude-slides it is Ask the agent to fix on a slide the editor can't show. In claude-design it is the same button on a canvas whose project/canvas.json can't be read, or on artboards with errors. In claude-design-system they are Sync from GitHub, Build from connectors and Create design system on an empty system, Fix with the agent on a contrast warning, Finish setting up, and Ask the agent about this file on an upload it refused. In claude-data they are Ask the agent on a part of the page you selected, Ask the agent to fix (or to limit) on queries that can't be shared, asking why sources failed, fixing a page that failed, Make editable, and a design-system change. With no agent listening, the slides, canvas and design-system buttons still copy their request for the chat, and the dashboard says the agent isn't listening.
• An empty deck, canvas, film, dashboard or board shows when the agent is at work on it, while the agent is on a comment or request it took, as on claude.ai.
• claude-motion: text style controls. In Tweaks in the side panel, a line of text has Weight (bold, italic), Size, Spacing (line and letter spacing) and Color, plus Align for text on several lines. A change shows on the film at once, is kept in modules/retouched.json and is in the frames stills makes. A row's Reset takes back that change, and Reset all changes takes back every change made by hand.
• claude-whiteboard: videos, a PNG export, cursor chat and following. Drop a video (mp4, mov or webm, up to 20 MB) on the board: it is uploaded and plays over its still. The agent may move or resize it but adds none, and upload still takes images only. The owner's Share panel has a PNG export of the board in its Export tab, before the whole folder. Everyone on a share link has cursor chat and can follow one another's view, Follow me included. Pages open on the same board now pass the board's room events to each other, so an upload in progress shows in the other people's pages too. Only asking the board's own assistant from cursor chat stays off: it needs claude.ai's model.
• claude-slides: video on a slide. A video added by link plays in its provider's player: YouTube's no-cookie player opens on a click, and Vimeo, Loom and Drive players are let in too. upload takes a clip for a slide (mp4 or webm, up to 20 MB), which the slide names with data-video on its still.

Fixes

• claude-motion: after any change made by hand (a scene's length, a retyped line, a color), opening Tweaks in the side panel turned the whole page into "This artifact couldn't be displayed". The Reset all changes button at the foot of the tab read something that was never set. The button now shows and works.
• claude-slides: the script in a live embed (<x-embed>) didn't run in the editor, because the editor page's script policy blocked it. It runs now. The embed is still sandboxed and still has no network.
• claude-whiteboard: a commenter on a share link was taken for a viewer. Follow me told them view-only access can't ask others to follow, and nobody could follow them. A commenter now comments and follows like everyone else.
• claude-whiteboard: a face sticker by a guest who had left read "Someone's sticker" once the page was reloaded. It now keeps the guest's name, unless their link was reset or they were removed.
• claude-data: comment mode left Edit, Design and Sources on. Now, as on claude.ai, they are off while you comment, and the Sources panel closes.
• claude-design-system: Download in the Files panel said downloads weren't available. It now saves the file.
• claude-design: an artboard copied from claude.ai whose name has spaces (or other characters the local server can't write) wasn't saved when pasted. It is now saved under a name the server can write, with the old name as its title, and another pasted artboard that imports it follows the new name. The context menu's Paste now reads images you copied. When the browser won't let it read, it pastes the editor's own copy.

Documentation

• The SKILL.md files of claude-slides, claude-design, claude-design-system and claude-data list the buttons that send the agent a request and say how one arrives. claude-whiteboard's says that videos on the board are people's, and what still doesn't run.

v2.15.2

Features

• claude-design: see a page at another screen width, and save the whole page at that width as a PNG. Open an artboard full window (the toolbar's Play, then the artboard) and pick Fill: next to Fill/Fit is a width control. It follows the window, or lays the page out at Phone 390, Tablet 768, 1024, 1280, 1440 or 1920 px (and the artboard's own width when it is none of these), centred, and scaled down to fit when wider than the window, with the scale shown beside the width. Whenever that width is not the artboard's own, the artboard's Export offers PNG 1×, this width and PNG 2×, this width: the whole page laid out at that width, as tall as its content, saved as <name>-<width>w@<scale>x.png. SKILL.md tells the agent to point users there when they ask for a page at a screen width or a full-page screenshot (requested by @hh0592821, #42).

v2.15.1

Fixes

• DeepSeek Harness: the Artifacts page showed one artifact to a row, each card as wide as the page. Since 2.14.1 a style meant for one card's drawing also applied to the whole list. The list is a grid of cards again.
• DeepSeek Harness: a plugin that was added from a folder and later uninstalled got no tile under Make something new, and the page could not install it. Its files stay in the profile after the uninstall, and the page took them for a plugin that is installed and switched off. The page now goes by what the app's plugin manager counts as installed, so the plugin gets its Not installed tile and installs from there.

v2.15.0

Features

• claude-motion follows claude.ai's 2026-11-10 update. A film that has a narration.js but no voice track of its own gets a Voiceover section under Tweaks in the side panel and a voiceover lane under the timeline. Change a line and click Record changes: the edit reaches the agent as a request (comments --wait prints it under requests; with no agent listening, the text is there to copy). A film with a voice track of its own (modules/voice.json) keeps its Script tab and lanes. write-db now refuses to write the records the page keeps of what you set by hand (modules/retyped.json, recolored.json, recut.json, refaced.json, retouched.json, fetched-fonts.json) and says PAGE_ONLY.
• claude-design-system follows claude.ai's 2026-11-10 update. Edits are saved as soon as the page is hidden (you switch tabs, say) instead of waiting for the autosave. A toast you close hands the focus back to where it was, and F6 reaches it.
• claude-slides follows claude.ai's 2026-11-10 update. A slide that can't be read, or whose markup may be partly hidden, says so on the slide and in the inspector, with Ask the agent to fix (with no conversation listening it copies a prompt to paste). A selected SVG image can't be set to Cover or Hug, and the inspector says why: it would change when the presentation reopens. While presenting, a swipe that starts on the page-number strip no longer turns the slide. A deck.json with a format this version doesn't know opens read-only, with a banner that it was saved by a newer version of Slides, and is not rewritten. Tables are saved with a wrapper for each row and with table, row and cell roles, so columns line up and screen readers read them as tables.
• Every editor draws Markdown in comments. Markdown in a comment or an agent's reply (bold, code, lists, quotes and so on) is drawn, not shown as characters; the All comments list keeps each row to flat text. The new-comment card shows a focus ring while you type in it.
• DeepSeek Harness: the Artifacts page's type tiles follow claude.ai's "Make something new" row. Each type is a tile with a pastel wash and a chalk doodle (dark mode swaps wash and ink), with redrawn type icons and a Beta badge on Data, Whiteboard and Motion. Rest the pointer on a tile, or Tab to it, and a card shows the type's name, what it makes and whether it is installed, with a button to start one, install it, or see the install. Escape closes the card; touch skips it.

Fixes

• claude-slides: its third-party licenses notice now carries the OfficeGen attribution that pptxgenjs's source header names.

Documentation

• claude-motion's SKILL.md and reference follow the same update. The agent ends with the film's url. A film for a loop, a GIF or a lobby screen has no score unless one is asked for. A request that says both what the film is for and how it should look skips the questions form. Plain films start from Inter and Space Grotesk. by-hand.md covers the page's records of lines retyped and colors set by hand.

v2.14.1

Features

• DeepSeek Harness: open your work from another device. When Harness runs on one computer and you use it from another (dsh web on a server, opened in a browser on your laptop or phone over the same network or a tailnet), opening an item from Artifacts, or View thread on a comment card, gave a http://127.0.0.1:… link that only that computer can open. The plugin now gives the device you are on a link of its own, at the address your browser reaches Harness by, and opens it in a new tab. Through it you are the editor's owner, as on that computer: you edit, and Send to Agent is there. Behind a proxy or a domain of your own, set BAOYU_DESIGN_REMOTE_HOST to the address other devices reach that computer at; off turns this off. Every editor's open takes --remote[=<host>] to print that owner's link (--remote=off closes it), and share <folder> --reset owner replaces its key. See Use it from another device (requested by @akaDeyve, #40).
• DeepSeek Harness: dsh-claude-dashboard updates to dsh-claude-data. The Artifacts page names the type Data, and offers an installed dsh-claude-dashboard as an update, claude-dashboard → claude-data: Update installs dsh-claude-data, switched on or off as the old one was, then uninstalls dsh-claude-dashboard. A preset that names dsh-claude-dashboard keeps it, and the page says what to change. 2.14.0 said dsh-claude-dashboard would get no more updates; instead it keeps getting releases with no skill in them. An Artifacts page from 2.13.0 or earlier offers that release as an update, and it brings the page that offers the move, so click Update once more when the move shows.

Fixes

• claude-design-system: a guest on a Share link could not add an asset or a font. The page stayed at Adding …, because a page served over plain http on the network lacks the browser's hashing that the upload used. It now hashes the file itself.
• DeepSeek Harness: uninstalling the plugin that was running the Artifacts page left the page at a 404 until the app restarted. Another installed plugin now takes it over at once.

v2.14.0

Features

• claude-dashboard is now claude-data. claude.ai renamed the type to Data, and the skill, its DeepSeek Harness plugin (dsh-claude-data) and its ChatGPT and Codex plugin (Baoyu Data) follow. Call it as /claude-data ($claude-data in Codex). A dashboard made with claude-dashboard opens in claude-data as it is, with the same folder; its address is now /data/<id>/ instead of /dashboard/<id>/. dsh-claude-dashboard and Baoyu Dashboard get no more updates: install dsh-claude-data (the Artifacts page offers it as a Not installed tile) or Baoyu Data, then remove the old one. On the Artifacts page, dashboards opened before the rename are listed with the new kind and still open, and comments sent from a dashboard reach the agent through dsh-claude-data (the copy of dsh-claude-data put on npm on its own before this release lacked both, because the Host plugin it came with did not know the new name).
• claude-data follows claude.ai's 2026-11-10 update. A dashboard's toolbar now has a Design button. It shows the look in use and a new Style section: chart colors, page color, font, corners and text size. A change shows on the page at once and is kept with Save; Reset all goes back to the defaults. A dataset can be a text file (format: "text"), which the page turns into rows with a loader. dash/meta takes the page's language (lang). check also reports what the open page found when it checked itself (pageCheck: passed, failed with what went wrong, or still loading) beside the page's runs.
• claude-design follows claude.ai's 2026-11-10 update. Pen, Line and Arrow have a Drawing options bar: four stroke weights, and solid, dashed or dotted lines. An artboard's export menu has All artboards (.png), a zip with one PNG per artboard on the page. When the canvas's saved layout is damaged, the canvas turns read-only with a banner and an Ask the agent to fix button, and comes back on its own once the file is fixed. A new reference, shaders.md, is how the agent makes a shader artboard (WebGL, or anything that moves like light, water, smoke or liquid).

Fixes

• claude-design: artboards ran React's development build (1.1 MB), which printed warnings such as Invalid DOM property vector-effect for drawn lines. They now run the production build (354 KB).
• claude-design: when two tabs changed the same artboard and you chose Keep both, the copy was put 80 px right of the original, where it could cover another artboard. It now goes in the first free spot right of that row.
• In every editor, Escape pressed while the keyboard is in the comments layer (on the comments button, say) now ends comment mode, as it does on the page.

Documentation

• claude-data's SKILL.md and reference follow the same update: what runs where (each reader's browser loads the datasets, and Sources shows where every number came from), what a page's code may write and when, and how check and runs report a page that went wrong.

v2.13.0

Features

• Whiteboard is released. claude-whiteboard lays out thinking on a board: a diagram, a flow, a map of ideas, a plan, a comparison or a drawn explanation. The agent writes it as elements (sections, blocks joined by arrows, text, tables, sticky notes, drawings and small working embeds) into the board's own store, and you watch them land in the local board editor. There you move and resize things, type, draw with the pen, and add sections, sticky notes, shapes and stickers from the toolbar, and the agent sees what you changed. Comment on an item, a spot or an area and send it to the agent, which changes the board and replies. Share opens the board to your local network with a link per role, everyone on a link sees the others' cursors and changes live, and its Export tab downloads the whole folder as a zip. History lists the versions. One server serves every board on the computer. Needs Node 22+. The board's own assistant on claude.ai (the Ask box, cursor chat, read-aloud and dictation) is not available locally; ask the agent in your conversation instead.
• DeepSeek Harness: the new dsh-claude-whiteboard plugin. Comments sent from a board reach the agent, and boards are listed on the Artifacts page with a picture of the board, where they can be renamed like the other kinds.
• ChatGPT and Codex: the new Baoyu Whiteboard plugin in the baoyu-design marketplace.

v2.12.3

Fixes

• claude-design: check now catches a canvas title placed too close above its row. An artboard's name strip keeps its size on screen while a title (title1) scales with the canvas, so a title 140 px above its row looked fine at 100% but covered the artboards' names once zoomed out below about 50%. SKILL.md already asked for at least 223 px. check now reports a title of 72 px or more that is closer than that to the artboards under it, or that sits over one: one issue per title, with the y to move it to. Text the user adds with the editor's text tool is left alone, wherever it is (contributed by @congee949, #38).

v2.12.2

Features

• claude-design: Export → HTML for Figma. Capturing the canvas, or the Standalone web page, with Figma's Chrome extension brought the artboards in as empty boxes, because each artboard is drawn inside a frame the extension does not look into. HTML for Figma (in the toolbar's Export, an artboard's Export and the Share panel's Export tab) saves <title>.figma.html: every artboard's rendered layers, placed where they are on the canvas, with their styles, fonts and images inside the file, and no frames and no scripts. Open it in Chrome and capture it with Figma's extension (Capture page, or Select element for one artboard). In an artboard's own Export it is that artboard alone. SKILL.md tells the agent how to answer when asked to bring a design into Figma (reported by @mufanmu, #36).

Fixes

• claude-design: PNG export no longer draws some icons as black or solid blocks. An SVG icon drawn with <use> lost the fill: none and stroke: currentColor its class gave it, and came out filled in black; it now takes the colour of the <use>. An icon cut out with a CSS mask-image (on an element or on ::before and ::after) was drawn as a solid block because the mask's file was not in the exported image; masks are now embedded like background images (reported by @hh0592821, #35).
• share gives out the Wi-Fi's address first on a computer with a VPN, a proxy or a hotspot. On Windows the links could start with the address of a proxy's virtual adapter (a sing-box TUN named falemon-tun, 172.19.0.1), which nobody else on the network can reach. Addresses of adapters named like a tunnel (falemon-tun, OpenVPN's TAP, Wintun, anything with vpn in its name), of point-to-point links, and of networks this computer makes itself (the Mobile Hotspot's 192.168.137.1, a virtual machine's or container's network) now come after the local network's. When a computer is on more than one network, share says that a link that does not open can be tried with another of the listed addresses (reported by @Bistard, #37).
• open no longer says the port is in use when a firewall or proxy filters ::1. A probe refused with EACCES or EPERM counted as an answer from something on the port, so open gave up on all 100 ports it tries.
• A page whose skill folder was moved, renamed or removed while its server ran now says so, and names the folder to the owner, instead of telling them to reinstall the skill. A guest is told to ask whoever shared it.

v2.12.1

Fixes

• A long comment thread no longer runs off the bottom of the window. The open card used to have no height limit, so the last replies and the reply box could be out of reach. In every editor it now stops at 60% of the window (never under 320px, and never taller than the room on the side it opens toward). The comments scroll inside it and the reply box stays below. After you reply, the thread follows to your reply; a long reply from the agent lines up at its top, and scrolling up stops the following (reported by @congee949, #34).
• claude-design and claude-slides opened on their own now keep the address bar on where you are. Under open or DeepSeek Harness's browser there is no claude.ai frame to do it, so switching pages or slides left the address as it was, a pasted #page-… or #slide-… link did nothing once the page had loaded, and Copy link to page and Copy link to slide could not copy. In claude-design the address now names the page, switching to another page adds a history entry (up to 20), and Back and Forward move between them. In claude-slides the address follows the current slide without filling the history. Both follow a changed address to its place, and the copy shortcuts copy the page's address with the place in it and say Link copied. On claude.ai nothing changes. This is the address part of #33, reported by @congee949; the previous/next page buttons and page keys it asks for are not in this release.