@marcelkargul The black fallback colours are the part that gets me. We keep the brand's colours, fonts and logo in a kit on the board, so a poster, an X banner and a sign-up screen all came out in the same brand.
@MediaKing Same goes for the creative side of each test. Poster, banner and sign-up screen for one idea, all in the same brand, without a design brief.
@cemtechconseil@neondatabase@mcpjams@manufact On the revocation path, the bit that got us was pending authorization codes. Disconnecting an app left an unexchanged code good for ten minutes, so the app reconnected itself. Delete the codes along with the tokens.
@chenzhangreal@yarotheslav Same idea on our side with better-auth. One OAuth provider covers the web sign-in and the MCP clients. We only allow the authorization_code and refresh_token grants, so every token belongs to a person who signed in.
@bayeschat@togethercompute@zainhas@parthsareen@ollama One more reason to keep a server's tool list static. Ours is 15 tools and one instructions string, the same for every user and session, nothing registered conditionally.
Asked one board for a poster, an X banner and an app sign-up screen.
All three came out in the same brand, because the board has a brand kit. When something was off, I said so and it got fixed in place.
@mcpdemolive Ours answers anything without a token with a 401 and where to sign in, tools/list included.
The less obvious one: an access token stays valid after someone disconnects the app. So every request also checks the app is still connected.
Building an MCP server that shows a card in Claude's chat?
Put the card's content in the result's _meta, not in the text.
Our card shows the slide Claude just made, and that's a whole page of HTML. In _meta only the card reads it. The model gets back the block id, where it landed and the board's name.
Past 100k characters we skip the preview.
@emadev01 For us it was previews. Our MCP server shows a little card of the slide it just made, and that card is a whole page of HTML.
It rides in the result's _meta, which only the card reads. The model just gets the block id, its position and size, and the board's name and link.
@maverickships Thanks Franco! Works with ChatGPT too: in its settings, add Flett as a custom connector and sign in. Then ask for the deck and it lands on the board.
@vibe_metrics@pshah_lab@X We don't count them yet, so no honest answer on which is more common. The report covers both: text past the edge or inside the margins, slides shrunk to fit, scripts or loads that failed. Logging which one fires most is a good idea.
@thisistonydang This is the fun version of what we build on whiteboards. What happens when two agents want the same pixel, last write wins? We let people and agents edit the same board and that question never fully goes away.
@KhaliqHussainnn Still up. No link, so here's the short version instead: Flett, a whiteboard Claude, ChatGPT or Cursor can draw on. Ask once, and a client deck lands on the board, ready to present. Honest take welcome.
@yigittbuilds Headings before and after is a good, cheap check. We'd run it on our eval set first, a fixed list of prompts we render and compare after every change. Real client decks only with the client's okay.
@owjuhl One thing worth building in early, from doing this on a whiteboard: the AI writing the view never sees it. Ours gets a report back after each render (content cut off, a script that failed, a load that broke) and fixes it before handing the view back.
@yigittbuilds Honestly, we haven't measured it. It's told what went wrong, not what to cut, so the AI decides. Checking that the key points survive is a good next test, thanks.
@sophylabshq Hi Ismail. Building Flett, a whiteboard any AI can draw on. The agency use: Claude builds the client deck on the board, you present it there, the client takes home a real PowerPoint.
@yigittbuilds Both, in that order. The slide scales itself down to fit, and if it has to go below 90%, the AI gets told the slide has too much on it and rewrites it. Shrinking further only hides the problem.
@paper Remote is also the only way into the web chats. Claude and ChatGPT on the web connect to a URL, not a local process, so that's where we started with our whiteboard's server. Good to see Paper there too.