Scaffold moves to Daytona
The code! command spins up a fresh sandbox, generates a full project, installs dependencies, builds it, and hands back a downloadable zip — all from one chat message. That sandbox layer has been migrated from E2B to the Daytona SDK. The move gives the scaffold pipeline a more predictable runtime to provision, install into, and tear down, which is exactly where a generate-then-build flow lives or dies.
A more reliable build pipeline
Migrating the sandbox was the opportunity to fix the long tail of "it generated but didn't build" cases. A batch of changes makes the output far more likely to compile on the first try:
- Guaranteed
package.jsonscripts — the project always has the scripts the build expects, even if the model omits one. - PATH-prefixed build commands and
npxfor tool scripts — the right binaries resolve inside the sandbox instead of failing with "command not found." - Forced devDependency install — build-time tooling is present, not skipped as it would be in a production install.
- Missing imports filled in, and a README always generated so every scaffold ships with usage notes.
- Robust zip download — proper error handling and a longer revoke window so the download link doesn't get pulled out from under the browser.
One frontend fix rode along: the scaffold preview card was being re-created on render, which reset its expanded/collapsed state. Hoisting the component to module scope keeps the preview stable while you read through the generated files.
See it run, right in the chat
A scaffold isn't just a folder of files you have to download to judge — it comes with a live preview that runs the generated project in a sandboxed frame inside the chat. Ask for an app, and you get a working, interactive preview you can click around in next to the source, without leaving the conversation or setting anything up locally. Browse the files, try the result, then grab the zip if you like it.
Pick your microphone
Voice chat now lets you choose which audio input it uses. Right-click (or long-press on mobile) the mic button and a device menu lists every microphone the browser can see. A few rough edges came with it and were smoothed in the same pass: the menu positions correctly, shows regardless of how many devices are attached, and a long-press on mobile opens the picker without accidentally starting a recording.
Built to be understood by AI crawlers
Search and AI assistants increasingly read sites through structured, machine-friendly files rather than scraping rendered HTML. GroupGPT now publishes a discoverability set — including an llms.txt summary and structured metadata — so when a model or crawler looks up "AI group chat," it gets an accurate, first-party description of what GroupGPT is and does instead of guessing from page chrome.
Why it matters
Scaffold is one of GroupGPT's most ambitious features — describe an app, get a built, downloadable project back — and its value is entirely about the success rate. Moving to a steadier sandbox and closing the common build failures makes the difference between a fun demo and a tool you'd actually use. The mic picker and crawler files are smaller, but they round out the same theme this release: make the powerful parts of the product dependable.