Open-source, privacy-respecting multi-model AI with the CrowClaw visual identity. {CROW-GODMOD3:ENABLED}
Crow-GodMod3 is an open-source research tool for studying LLM output steering and safety evaluation.
By using this tool, you agree that:
This is an experimental research preview. Models are accessed through the provider or local endpoint you configure. Crow-GodMod3 does not publish application telemetry; selected providers and hosting infrastructure have their own policies. See the terms for details.
When enabled, this prompt is prepended as a system message to every API call. Use it for persona instructions, style guidance, custom rules, or anything you want the model to follow.
Configure parameters for each mode. Select modes from the dropdown in the main interface.
Streams tokens live as they arrive from a single model, instead of waiting for the complete response.
Shows the first scored response immediately, then morphs the display as faster models finish and a higher-scoring leader emerges. Each upgrade must beat the current leader by the min-delta threshold.
All jailbreak templates race in parallel instead of early-exiting on the first non-refusal. The best-scoring template is served immediately, then progressively upgraded as higher-scoring templates complete. The winner is then handed to the Liquid refinement loop for iterative polishing.
Parseltongue operates on the input side (obfuscating trigger words before they reach the model), while Liquid operates on the output side (upgrading responses as they arrive). They are fully independent and stack: Parseltongue encodes your prompt, the model responds, and Liquid delivers the result with live morphing.
localhost or 127.0.0.1. Docker Model Runner uses the nested /engines/v1 path.
/models. There is no model-count limit. Each mode’s Automatic pool is configured independently under Strategies; a header pin still runs one exact model.
. Your browser may also ask to allow local-network access; approve that prompt for discovery and chat.