A single-file chat UI that races jailbroken models in parallel

What it does G0DM0D3 is a chat interface that aggregates access to over 50 large language models through OpenRouter, wrapping them in a browser-based UI contained entirely in one index.html file. It runs without a build step or server-side dependencies, storing your API key and chat history strictly in browser localStorage. The project is explicitly aimed at red teaming, cognition research, and what it calls “liberated AI interaction” — essentially probing how models behave when pushed outside standard conversational boundaries.

The interesting bit Rather than treating model refusals as final, G0DM0D3 weaponizes them: its GODMODE CLASSIC mode fires five distinct model-and-jailbreak-prompt combos in parallel to surface the least-restricted response, while ULTRAPLINIAN scales this up to as many as 51 models with a composite scoring engine. There is even a dedicated Parseltongue module that obfuscates trigger words via leetspeak, Braille, Morse code, and Unicode substitution to test model robustness — all inside a vanilla HTML file that weighs nothing and deploys anywhere.

Key highlights

• Parallel model racing: GODMODE CLASSIC pairs specific models with battle-tested jailbreak strategies (boundary inversion, refusal inversion, l33t formatting) and returns the winner; ULTRAPLINIAN expands this to tiered fleets of 10–51 models scored on a 100-point metric.
• Red-team perturbation engine: Parseltongue detects trigger words and applies 33 obfuscation techniques across three intensity levels, including phonetic substitution and bubble text.
• Adaptive sampling: AutoTune classifies queries into context types and automatically adjusts sampling parameters (temperature, top_p, etc.), refining its choices via an EMA loop fed by thumbs-up/down feedback.
• Output normalization: STM Modules post-process responses in real time to strip hedging language, remove preambles, or inject curiosity prompts.
• Radical self-hosting: The entire client is one static index.html file; there is no login, no cookie, and telemetry is opt-out with no message content collected.

Caveats

• The optional API server can publish full chat logs to a public Hugging Face dataset if you opt in; the README warns that automatic PII scrubbing is not guaranteed and the data may be redistributed.
• All chat history lives in browser localStorage only — clear your cache or switch devices and your conversations vanish without recovery.
• The README lists “GPT-5” among supported models, which appears premature or aspirational at the time of writing.

Verdict This is for security researchers, prompt engineers, and developers who want to stress-test model boundaries without spinning up infrastructure. If you are looking for a standard, guardrail-compliant corporate chatbot, this is explicitly the wrong tool.