Changelog

• Multiple Telegram bots — beyond the Default bot you can now add extra bots in Settings → Telegram (each with its own BotFather token), and rename or remove them. The Default bot still handles everything automatic; extra bots are for connecting terminals.
• Pick a bot per terminal — the bot icon in a board’s terminal toolbar opens a menu (Default, Bot 1, …) so each terminal session can connect to a different bot. Run several terminals at once, each driven from its own Telegram chat. With only the Default bot configured, the icon stays a simple on/off toggle.
• Agent media goes to the connected bot — when the agent sends media from a terminal (“send me that image”), it now reaches only the bot that terminal is connected to, instead of being shared globally.

• New “Folder” card — drag or paste a folder onto the canvas to drop a Folder card. It shows a thumbnail of the first image found in the folder (or a folder icon) and an Open folder button that reveals it in Finder/Explorer.
• Import media from a folder — the Folder card’s Import button opens a picker listing all media in the folder, where you can select items to import onto the canvas (or remove files from the folder to Trash).
• Crash recovery — if the display process exits unexpectedly the app now auto-reloads (with a fallback dialog if it keeps happening) instead of leaving a black window, and a top-level error boundary keeps a single failed component from blanking the whole window.

• Fixed: Intel (x86_64) Mac build failed to launch (“Backend failed to start”) — the Intel build was shipping an Apple-Silicon backend; it now builds and packages a matching x86_64 backend under Rosetta, so it launches correctly.
• Clearer backend-startup errors — when the local backend can’t come up, the dialog now reports the real cause (port already in use, an early crash, the last backend output) instead of a generic timeout, and fails fast instead of hanging.
• Configurable backend port — force a specific port with the GREENLIGHT_BACKEND_PORT environment variable for machines where the default range is occupied.

• MotionFast-style render progress — a clear progress bar overlays the preview during export, with a reminder to keep the editor open while it renders.
• Exports land on the board — the rendered video is uploaded to the current board as a video card (with a notification sound), in addition to the regular file download.

• Export frame-rate selector — choose 24 / 30 / 60 / 120 fps next to the format dropdown (it was hardcoded to 30 fps).
• Unsnapped channels survive reload — the per-channel “unsnap layers” state and each layer’s free position now persist across reload, project switching and JSON template import, instead of snapping back to sequential.

• Font Size now works — the camera’s zoom-to-fit was cancelling out font size exactly; framing now scales with the font size you choose.
• Video length retimes correctly — adding or removing words now updates the timeline and the exported video duration (it used to be captured once at template load).
• Speed slider step refined for finer control.

New premium image and video models from GeminiGen.

• GPT Image 2 — text-to-image and edit, with quality modes (low / medium / high), resolutions up to 12K, and 8 aspect ratios including 21:9, 3:2 and 2:3.
• Seedance 2 — a new first/last-frame video model with a Fast/Pro selector, 4–15s durations and 6 aspect ratios.
• Seedance 2 Omni updated — up to 4 reference images, new reference video and audio inputs, and the full Fast / Pro / VIP mode lineup.

• Fixed flashing of HTML and PDF cards while scrolling the board — big HTML pages went blank mid-scroll and popped back in after a reload. Iframe-backed cards (HTML, PDF) are no longer unmounted when they scroll off-screen, so panning past them is seamless.

Stops Windows Defender’s ML heuristics from false-flagging the installer (Microsoft’s own deep analysis returns “No malware detected”).

• Every executable inside the app is now EV-signed — the app, backend, yt-dlp, ffmpeg/ffprobe and the Telegram bot, not just the installer — removing the unsigned-inner-binary pattern that tripped the heuristics.
• Removed the legacy winpty binaries — the terminal uses Windows’ built-in ConPTY on Windows 10 1809 and newer.
• Updated bundled downloaders — yt-dlp and pytubefix refreshed to current releases.

A big rollup of the whole 2.8.x series. The headline additions:

New

• Voice generation — a Voice button in the bottom toolbar with two Chatterbox models: Text-to-Speech (optional reference voice, plus exaggeration / temperature / CFG controls) and Speech-to-Speech. You can record your own voice from the mic, and results land on the board as playable audio cards.
• Draw on images — annotate photos directly on the canvas with drawing tools; the marked-up image is saved back to the board.
• HTML editor Design System — manage root-level design tokens (create, edit and delete variables live), pick board / media-library / uploaded media for image & video sources, set device widths for HTML cards (Mobile / Tablet / Laptop / Desktop) and fit-height for Markdown cards.
• New AI models — Luma Ray 3.2 image-to-video, Seedance 2.0 Spicy image-to-video with native audio, and accurate resolution- and duration-aware price estimates for every per-second WaveSpeed video model.
• Boards & files — low-res WebP previews keep heavy photo boards responsive, Figma .fig files drop onto boards as preview tiles, clean human-readable file/board naming, and you can share a board straight from the boards list.

Fixed

• A round of HTML-editor reliability fixes — file corruption on text edits, vanishing subselector styles, and Design System variables that wouldn’t apply or delete.
• An oversized editor media picker, an unstyled board-name menu, broken thumbnails in Recent Generations, and the screenshot tool failing in the packaged app.