Back to Playground
2026 • MANIFESTO

KIT OS

Beyond the Chatbox: The Era of Encapsulated Micro-LLM Apps.

Launch App
Gemini 2.0 Generative UI Agentic Core
Use arrow keys or buttons below to navigate
PART I: THE TRAP 02 / 06

The "Chat" Interface Trap

Modern chat interfaces are ephemeral buckets of lost potential.

Prompt Drift

You get a great result once, but can't repeat it exactly 3 days later because you forgot the nuance.

The Search Problem

Searching for "that one prompt", "that one output", or "that one AI file" in a linear chat history is impossible.

No Persistence

Chat implies transience. Real work requires state, history, and versioning.

PART II: THE SOLUTION 03 / 06

Encapsulated Micro-Apps

Don't "chat" with AI. Use tools powered by AI.

The Capsule Config

{
  "input_schema": { ... },
  "system_prompt": "...",
  "output_renderer": "..."
}

Every app is just a database entry defining this triad.

The Toolmaker

A meta-agent that writes these configs. You describe the workflow, it builds the persistent tool.

PART III: ENGINE 04 / 06

Multimodal Native

We leverage the latest model breakthroughs to power these capsules.

⚡️

Gemini 2.0 Flash Lite

Low latency, high vision capabilities. The "Eyes" of the system.

🧠

Mistral / Devstral

Strict logic and code generation. The "Cortex" of the system.

PART IV: INTERFACE 05 / 06

Generative UI

The interface should adapt to the data, not the other way around.

Challenge Solution
Ambiguity Strict Schemas (Zod)
"Hallucinated" Widgets Component Fallbacks
Unpredictability Validation Layers & Dynamic Rendering

THE VISION

A living OS where you build your own tools in seconds, and they persist forever.

Launch Kit
© 2026 Kit OS

Detailed Product Specification

Full writeup below

BACK TO PLAYGROUND

Kit OS Manifesto

STATUS: ALPHA • PLATFORM: WEB • YEAR: 2026
Micro-Apps Toolmaker GenUI
PART I: THE PROBLEM

The "Chat" Trap

We are currently stuck in the "CMD.EXE" era of AI. The chatbox is a command line interface that pretends to be a conversation, and it fails at repeatability.

1. Prompt Drift

You craft the perfect prompt on Tuesday. On Friday, you try to recreate it, but you forget one specific instruction. The output changes. Reliability collapses.

2. The Search Black Hole

Trying to find "that one CodePen I generated three weeks ago" in a linear chat history is a nightmare. Chat logs are not a file system.

3. No State

Chat is ephemeral. Real software needs state. It needs to know "this is the current draft," not "here is a wall of text from 5 minutes ago."

PART II: THE SOLUTION

Encapsulated Micro-LLM Apps

The solution is to stop treating AI as a conversational partner and start treating it as a **component engine**. We build "Capsules."

The Capsule Triad

  • Input Schema: Strictly typed arguments (e.g., "Code snippet", "Target Language").
  • System Prompt: The frozen, perfected instructions that never drift.
  • Output Renderer: A UI component specifically designed to view the result (e.g., a Syntax Highlighter, a Chart, a Table).

By saving this triad to a database, we create a Tool. This tool is repeatable, searchable, and shareable.

The Toolmaker

You don't need to code these tools manually. The **Toolmaker** agent accepts a natural language description ("Make a tool that converts Python to Rust and explains the memory safety differences") and generates the configuration instantly.

PART III: TECHNICAL STACK

Multimodal & Generative UI

To make this work, we need speed and strictness.

Gemini 2.0 Flash Lite

We use this for the "Eyes." Its multimodal native capabilities allow us to drop screenshots or mockups into a tool content.

Mistral / Devstral

We use this for the "Hands." Its superior code generation logic ensures the tools work as expected.

Generative UI (GenUI)

Standard UI is static. Chat UI is text. GenUI is the middle ground: UI that spawns based on the data.

The Risk: Hallucinated interfaces (AI trying to render a component that doesn't exist).

The Fix: We implementation strict Zod schemas that force the AI to choose from a registry of valid components (Charts, Tables, Markdown, CodeBlocks). If it tries to invent a widget, the validation layer catches it and falls back to raw text.