Test and compare AI model responses with an open-source chat interface
Imagine asking a single question and having multiple AI models answer simultaneously, allowing you to select the best result. This tool lets you compare multiple AI models on one screen and observe their reactions by applying specific prompt strategies. It is not just a chat app but a research tool for testing the limits and possibilities of AI.
Model competition and evaluation engine
The core of this tool is its ability to run multiple models simultaneously. GODMODE CLASSIC runs five verified model and prompt combinations in parallel to find the best answer. ULTRAPLINIAN uses a more complex evaluation engine, querying between 12 and 60 models at once and scoring them with a composite metric out of 100 points to determine the winner. This process is based on OpenRouter's model list, and you can also add Venice or local models.
Input transformation for red teaming
Parseltongue is an input transformation engine for testing the robustness of AI models. It applies 33 transformation techniques at three intensity levels, allowing you to study how models react to specific trigger words. These techniques include leetspeak, braille, Morse code, and Unicode substitution. Additionally, AutoTune classifies the context of a question into 20 types and automatically adjusts sampling parameters like temperature or top_p accordingly.
Local model integration and privacy
This tool connects to local models running on your own hardware via Ollama, LM Studio, llama.cpp, and vLLM. Enabling Local-only mode blocks calls to external services like OpenRouter or Venice and automatically disables app telemetry. Conversation history and settings are stored only in the browser's local storage, with no account system or cloud synchronization. Therefore, deleting browser data permanently erases conversation history, making it important to back up using the export function.
Pre-use considerations
This tool is distributed under the AGPL-3.0 license, which requires you to disclose the source code if you provide a modified program over a network. Telemetry is enabled by default but can be disabled with No-Log mode. However, the logging policies of the model providers themselves or network logs from hosting infrastructure are not controlled by this tool's settings. For complete privacy, it is safest to use local models, enable Local-only mode, and self-host.
