Developer Stephen Dale has launched NoFlattery, a local-first, single-file web application designed to help users make better decisions by utilizing a council of adversarial AI agents. Unlike standard AI assistants optimized to validate user input, NoFlattery employs 2–4 agents programmed with different reasoning biases to debate a user's decision. The application delivers a structured Decision Record rather than a standard chat transcript, providing a verdict, reasoning, main risk assessment, potential conditions for changing the decision, and actionable next steps. The tool is intended for high-stakes decision-making in areas such as product strategy, pricing, hiring, or technical architecture.

The project is built as a single-file Single Page Application (SPA) using React 19, TypeScript, and Zustand for state management. To ensure local-first functionality, the application utilizes Dexie to store conversation history and API keys directly in the user's browser using IndexedDB. This architecture eliminates server-side data storage, telemetry, and complex build orchestration, allowing the app to be deployed as a single index.html file to platforms like Cloudflare Pages. Users are required to provide their own API keys (BYOK) for providers including OpenAI, Anthropic, OpenRouter, or Ollama, ensuring data privacy.

The core engine operates on a deterministic, turn-based model where each agent speaks in a fixed order based on defined biases. The system supports three specific discussion modes implemented through prompt injections: Adversarial mode, which encourages agents to challenge ideas; Audit mode, which prompts agents to identify flaws; and Vote mode, which forces a definitive conclusion. According to Dale, avoiding complex, model-based orchestration was essential to maintaining system reliability and debuggability. By keeping data entirely within the browser, the developer intends to establish user trust, positioning the product as a private, reliable alternative to traditional, chat-focused AI tools.

Developer Stephen Dale has launched NoFlattery, a local-first, single-file web application designed to help users make better decisions by utilizing a council of adversarial AI agents. Unlike standard AI assistants optimized to validate user input, NoFlattery employs 2–4 agents programmed with different reasoning biases to debate a user's decision. The application delivers a structured Decision Record rather than a standard chat transcript, providing a verdict, reasoning, main risk assessment, potential conditions for changing the decision, and actionable next steps. The tool is intended for high-stakes decision-making in areas such as product strategy, pricing, hiring, or technical architecture.

The project is built as a single-file Single Page Application (SPA) using React 19, TypeScript, and Zustand for state management. To ensure local-first functionality, the application utilizes Dexie to store conversation history and API keys directly in the user's browser using IndexedDB. This architecture eliminates server-side data storage, telemetry, and complex build orchestration, allowing the app to be deployed as a single index.html file to platforms like Cloudflare Pages. Users are required to provide their own API keys (BYOK) for providers including OpenAI, Anthropic, OpenRouter, or Ollama, ensuring data privacy.

The core engine operates on a deterministic, turn-based model where each agent speaks in a fixed order based on defined biases. The system supports three specific discussion modes implemented through prompt injections: Adversarial mode, which encourages agents to challenge ideas; Audit mode, which prompts agents to identify flaws; and Vote mode, which forces a definitive conclusion. According to Dale, avoiding complex, model-based orchestration was essential to maintaining system reliability and debuggability. By keeping data entirely within the browser, the developer intends to establish user trust, positioning the product as a private, reliable alternative to traditional, chat-focused AI tools.