App Builder
Describe an app or show a reference picture, and AI writes a working web app you can preview, edit and publish to your own RunApps account.
+25% creator fee
Turn an idea into a working web app by chatting with AI. Describe what you want, or show it a reference picture, and it writes real HTML, CSS and JavaScript files you can preview, edit and publish to your own RunApps account.
Features
- Chat with streaming Markdown replies; the work of each answer (files read and changed, previews, screenshots) folds into one line you can open, with a card per step.
- AI tests what it builds in the live preview. It reads the page (text and every control in page order, with new and covered controls marked), uses it like a person — click, double-click, right-click, hover, type, press keys and shortcuts, pick options, tick boxes, drag (pointer drags and HTML drag and drop), scroll, upload project files, wait for things to appear or disappear — runs JavaScript in the page to check state, and switches between desktop, tablet (768 px), phone (390 px, touch) or any custom size. Console output, errors and dialogs come back with each step.
- Watch the AI work. While it operates the preview, the preview gets a thin accent border and an "AI is using the preview" label, and the AI's own cursor glides to each target, outlines it, then clicks with a ripple and types at a readable pace. Your mouse is held off the preview during each step so it cannot interfere; all of it goes when the task ends.
- AI asks one to three questions when a requirement is unclear, one at a time in the chat: pick an option (nothing is chosen for you) or type your own answer. Questions and answers are kept with the project.
- Attach pictures or files with the paperclip, by dropping or pasting; they are saved in the project, and models that see pictures receive them.
- Edit code with a file selector, visible save controls, keyboard shortcuts and automatic saving. Search, rename, upload and export files, compare versions, or restore a snapshot.
- Sign in to automatically restore and synchronize projects, code, conversations and file versions. Accounts have separate workspaces. Offline changes retry later; conflicts keep your work available instead of silently overwriting it.
- Models that see pictures can capture and inspect the current preview without reloading it; the screenshot shows in the chat. Image support is detected from the live model catalog.
- Preview multiple pages on desktop, tablet and phone layouts, inspect console errors, or select an element to ask AI to change it.
- Choose a task list, portfolio, focus timer or AI assistant template. Generated apps can use the RunApps SDK.
- AI can query the live model catalog and inspect a model’s parameter schema before implementing an integration, without hardcoding today’s model list.
- The chat's settings (the gear) choose the model (those that see pictures are marked), thinking level, turns per run (100 by default; continue when it is reached), reply length and tool permissions, per account. The ring around the send button shows how full the context is; hover it for cost, tokens and calls. Older conversation is summarised automatically as the model's context fills.
- Connect browser-accessible Streamable HTTP MCP services. Every external tool call asks first, showing the server, the tool and its arguments.
- Publish generated apps under your own account, unlisted or submitted to the library. Publishing authorization is separate from app sign-in.
Good to know
AI usage is charged to your RunApps balance. A request may use several model calls; context summarization is also billed. Opening templates, manual editing and loading a preview do not call a model. AI actions inside previews require confirmation and use the current Builder account.
The AI's preview actions are simulated events inside the sandboxed preview: they reach the app's own code but not the browser itself, so CSS :hover styles, native file pickers and real page navigation are not exercised. Confirm and alert dialogs are answered automatically while the AI is acting.
Apps are static HTML, CSS, JavaScript (including ES modules), SVG and images. There is no npm build, JSX/TypeScript compilation or backend process. A project supports up to 80 files and 10 MB (600 KB per text file and 2 MB per image), with 12 file snapshots.
Preferences stay in the current browser. Clearing browser data removes unsynchronized local work; changes waiting to sync are kept in the browser and upload the next time the page is opened.
Conversation and memory
Clear (in the chat header) removes the conversation, its summary and spending while keeping code, file versions and project memory. Project memory (the notebook in the chat header) holds up to 8,000 characters of durable notes; edit it there, or ask AI to remember or forget a requirement. It syncs with the project and is included in every request.
Test AI before publishing
The isolated preview can use your current Builder account for chat and image generation. Each paid request shows its project, account, method and parameters for approval. Credentials and Builder cloud files stay outside the preview. Reloading a preview or switching accounts cancels pending requests; already submitted jobs may still be billed. Preview data stays separate from a published app. Audio, video and cloud-file APIs are not currently relayed.
- Read sourced SDK documentation on demand and create private prompt files that are included in exports and publication. Private prompt references must be tested in the generated app after publication; the isolated Builder preview does not resolve them.
Preview screenshots use DOM rendering. Cross-origin assets, video, web fonts and some visual effects may be missing or differ from the browser display.
The AI cursor and preview mask draw on the design of alibaba/page-agent (MIT).