--- title: "Code Mode: From One Prompt to a Running App" slug: "code-mode" date: "2026-08-12" tag: "Application" summary: "Code Mode is the coding agent built into the Agentic OS chat box. One prompt scaffolds a Next.js app from the Agentic Vibe starter, drops in pre-built ibl.ai components, installs dependencies, and runs it locally — with every tool call gated behind an explicit approval." author: "ibl.ai Engineering" repo: "iblai/vibe" linkedin: | "Vibe coding" usually means a chat window in one tab and the thing you are building in another, plus a running argument about which files the model is allowed to touch. A new walkthrough of Agentic OS shows Code Mode, the coding agent sitting inside the chat box itself. One prompt — "build a todo list website" — and it scaffolds a Next.js app from the Agentic Vibe starter template, drops in pre-built ibl.ai components, installs dependencies, and runs it locally for preview. The frame worth pausing on is the one where it stops. Partway through the run the agent asks: "Code needs your permission — READ. Allow or Deny." An agent that writes files into a workspace is precisely where deny-by-default belongs, and here it is a setting rather than a promise. The speed is not the model being clever. It is the toolkit underneath: the scaffold already carries SSO, AI chat, profiles, and analytics, and 30+ skills teach the agent how to build against the platform. The agent is composing pieces the organization already owns instead of inventing a stack from scratch. Which is the part that matters after the demo ends. On ibl.ai you own all the code and the data — the toolkit is MIT-licensed on GitHub, it runs model-agnostic against any LLM you point it at, and one codebase ships to web, desktop, and mobile. You keep the app. #iblai #AgenticAI #VibeCoding #AI #DeveloperTools --- **Code Mode** is the coding agent built into the Agentic OS chat box. The toggle sits in the same composer as Cowork, Canvas, Prompts, and Memory — there is no separate IDE to open, no second tool to authenticate, and no context to re-explain to a different assistant. A new walkthrough takes it from an empty workspace to a running application in a single conversation. ## What does Code Mode do with a single prompt?

Code Mode scaffolding and running an app from a single prompt.

The prompt is four words: `build a todo list website`, typed into the chat box with the **Code** toggle enabled. The agent then plans out loud before it touches anything: > "I'll scaffold this as an ibl.ai Next.js app first, then add the todo list UI within that app and run it locally for preview. The workspace is empty apart from `.git`, so I'm materializing the vibe-starter template and installing dependencies." From there the run is: scaffold from the Agentic Vibe starter template, request permission, install dependencies, start the dev server. The finished result is **Focus Board** at `localhost:3000` — ibl.ai logo and navigation, Home / Profile / Account routes, priority tags, Active and Done filters, and live counters — assembled from the pre-built ibl.ai components rather than written from nothing. A task is added live at the end of the video to show it working. ## What happens when the agent wants to touch your workspace? It stops and asks. Partway through the run a gate fires — **`Code needs your permission — READ`**, with **Allow** and **Deny** — and the agent waits. This is the frame worth pausing on, and it is easy to miss in a demo about speed. An agent that reads and writes files inside a workspace is exactly where deny-by-default belongs. Approval is not a courtesy dialog bolted on after the fact; it is the execution model, and it holds for every tool call, not only the first one. It pairs with the controls already documented under [Agent Settings: Sandbox](/docs/os/agent-settings/sandbox) — isolated execution, allowlisting, and credential injection — so an organization can decide what a coding agent may reach before anyone types a prompt, and still see each individual action before it happens. ## Why is it fast — the model, or the toolkit? The toolkit. **Agentic Vibe** is the layer underneath Code Mode, and it is doing most of the work the model does not have to. One command generates an application with SSO, AI chat, user profiles, and analytics already wired up, so the agent is not inventing authentication or a chat transport on the fly. Alongside it, **30+ Claude Code skills** teach the agent how to build against the platform — the same reusable-playbook mechanism we shipped as [Agent Skills](/updates/agent-skills) — which turns adding a feature into a conversation rather than a documentation scavenger hunt. The result is an agent composing building blocks the organization already owns. That is a different failure profile from a general-purpose model guessing at a stack: fewer invented APIs, fewer half-configured integrations, and a codebase that looks like every other application your team maintains. This is the same story as the [App CLI](/updates/app-cli-frontend-scaffolding), driven from a chat box instead of a terminal. The CLI scaffolds the same project from the command line; Code Mode does it inside the conversation, and then keeps going. ## What do you keep after the demo ends? The application. The toolkit is MIT-licensed on GitHub, and on ibl.ai you own all the code and the data — the generated project is ordinary Next.js source in your own repository, building with your toolchain and deploying to your infrastructure. It is model-agnostic across OpenAI, Gemini, Anthropic, and Llama, so switching models is a configuration change rather than a rewrite. And the same codebase ships to web, desktop, and mobile through Tauri v2. Agentic Vibe is documented on the [Agentic Vibe product page](/product/agentic-vibe), and the skills mechanism Code Mode relies on is covered in [Agent Settings: Skills](/docs/os/agent-settings/skills).