Over the past year, many SaaS products have added AI chatbots to answer questions and reduce support load. It helped initially, but most of these systems still live in a chat window with no awareness of what’s happening inside the application. They don’t know the current page, selected data, user permissions, or workflow state. Users end up repeating context the product already has. I recently came across an open-source Copilot SDK that approaches this differently by injecting live application state directly into the AI and letting it execute real frontend and backend functions instead of just responding with text. What it does: - Understands application state including current page, selected data, and user permissions - Executes backend and frontend functions instead of only responding with text - Delivers richer product experiences through generative UI such as tables, forms, and interactive buttons - Understands user workflow and intent based on in-product context - Maintains session context so interactions remain consistent Example: Instead of the AI asking "What do you need help with?", it understands the user context is viewing failed transactions from last week and can immediately offer to retry them, export the data, or investigate patterns. Technical details: - Works with React, Next.js, Vite (Vue & Angular coming soon) - LLM-agnostic (bring your own model) - State injection via context providers - Tool execution layer for safe function calling - Full data ownership (everything runs in your infrastructure)

Docs + examples: https://copilot-sdk.yourgpt.ai

Happy to answer technical questions about implementation, or specific use cases.