In October, we shared why we’re building NextLLMs. We wanted a place where a conversation could become something useful, with the files, tools, and decisions from that conversation close at hand. This month’s work makes that idea easier to see. You can ask for a small website, open the files it creates, edit them, and run a preview beside the chat. If you spot something you want to change, you can ask for a follow-up while keeping the project in view.
Here’s what we’ve been building since that first post. Some of the client updates below are still being prepared for wider distribution.
From a chat message to editable project files
Project Files gives each coding chat its own workspace. When you ask for a browser prototype, the assistant can create real files such as index.html, styles.css, and app.js. An example inside a code block stays in the conversation. It only becomes part of the workspace when it’s saved as a project file.
On desktop, the editor can sit beside the conversation or expand to full screen. You can browse folders, open multiple tabs, edit files, and download the project as a ZIP. When the assistant changes an existing file, the conversation can show the affected files and let you review the changes or undo them. If a file changes while you’re editing it, the workspace can ask you to review the newer version before saving over it.
Run opens an isolated preview of a browser project using the files in that chat. The file explorer folds away to make more room, and comes back when you stop the preview. We’ve also been working on clearer error messages. When an error can be traced to a project script, the preview can show the file and, where available, a line number. A browser extension can also throw an error inside the frame, so we’re improving how the preview distinguishes those reports from errors in the project itself.
A simple way to try the workflow is to ask for a small website with separate HTML and CSS files. Open Project Files, change a headline, run the preview, then ask the assistant to improve one section. The change review lets you see which files the follow-up touched.
Skills for work you repeat
Some tasks come with instructions you give every time. You might want code reviews to follow your team’s checklist, or product copy to use the same vocabulary and structure. A Skill gives you a place to save those instructions and their supporting material.
Under Settings → Skills, you can import a standalone SKILL.md, a skill folder, or a ZIP. Once enabled, a skill appears in the composer’s Skills menu and can be called with a slash command such as /review-code. The assistant can also select an eligible skill that matches the task. It loads the instructions and relevant reference material when needed, so a long checklist doesn’t need to be pasted into every conversation.
Skill packages can contain scripts as well as text. Compatibility varies by package, and execution depends on the sandbox and the permissions available for the task. Review a package before enabling it, especially if it came from someone else.
Auto Router and more useful feedback while you wait
Auto Router chooses an eligible model based on the task, the account’s plan, and the capabilities the request needs. A quick question, a coding task, and a request that needs tools can call for different models. You can still choose a model manually where the picker offers one. Model information in the picker helps explain what each option is intended for.
The Effort control still lets you choose how much reasoning to request where the selected model supports it. During longer requests, some providers can stream short progress summaries. When a provider doesn’t supply a summary, NextLLMs shows that the request is still pending. Tool activity and progress are easier to follow when a response uses several steps.
We’ve also worked on how the chat shows plan and credit information. When a subscription check is still in progress, the app can wait for the verified allowance instead of briefly showing the wrong plan limit. That matters when you’ve just signed in or when your subscription has recently changed.
Web images and generated media
When you ask for an image from the web, NextLLMs can search for candidates, check whether they fit the request, and show selected images with their source links. Author or licence information appears when available. You can follow the source to check the original context or the terms for reuse.
Image generation remains a separate workflow. We’ve also worked on restoring generated video after a reconnect and making it easier to play a finished clip from the conversation.
On Android, in VS Code, and on the desktop
On Android, recent builds add a skippable first-run guide, editable starter prompts, and app settings for language, reduced motion, and notifications. Launcher shortcuts provide quicker access to a new chat, voice input, or settings. Android release and notification availability depend on the installed build and rollout.
In VS Code, you can choose which local context to share and optionally connect a trusted project folder to a chat’s Project Files. The extension keeps the composer’s context chips focused on top-level files and folders. When a website edit reaches a connected folder, the sync workflow can ask for approval and flag conflicts with local edits. We’re also working on local Python workflows for connected Windows projects.
NextLLMs Code is a separate Windows desktop project we’re developing alongside the extension. It’s currently a development build. We’ll share availability details when it’s ready for a broader release.
For eligible accounts, we’ve also added a sign-in approval option that uses another signed-in device. The device requesting sign-in shows a code, and the approval appears on an open NextLLMs page on the other device.
Tell us where the workflow breaks
The most useful feedback often comes from a real task. If a preview fails, tell us which project file you ran, what you expected to see, and what happened after you clicked Run. If the issue is with a model, a skill, or a connected editor, include the option you selected and the steps leading up to the problem. A screenshot or a small example that reproduces the issue helps us narrow it down.
You can send a report under Settings → Support → Bug report. Leave out passwords, API keys, and private data. Thanks for building with NextLLMs and telling us what gets in the way.