← all repositories
elder-plinius/G0DM0D3

One HTML file, 55 models, zero guardrails

A single-file chat UI that races jailbreak prompts against multiple LLMs and scores which one "wins."

7.5k stars TypeScript Chat AssistantsLLMOps · Eval
G0DM0D3
Velocity · 7d
+100
★ / day
Trend
steady
star history

What it does

G0DM0D3 is a browser-based chat interface built as one self-contained index.html file. You bring an OpenRouter API key; it gives you access to 50+ models (Claude, GPT-4o, Gemini, Grok, etc.) with a twist: it deliberately runs multiple “jailbreak” prompt variants in parallel and picks the best response. There’s also a red-teaming mode that obfuscates inputs via leetspeak, braille, Morse code, and Unicode tricks to test model robustness.

The interesting bit

The whole application — UI, logic, styles, state management — lives in a single HTML file with no build step, no framework, no server required. Deployment is literally scp index.html to any static host. The “ULTRAPLINIAN” mode will fire your query at up to 51 models simultaneously, score responses on a 100-point composite metric, and crown a winner. It’s overkill as a product feature, but it’s a neat research primitive.

Key highlights

  • GODMODE CLASSIC: 5 hardcoded model+prompt combos race in parallel; “best response wins” (the README is vague on how “best” is determined)
  • Parseltongue: 33 input perturbation techniques across 3 tiers for red-teaming research
  • AutoTune: EMA-based online learning from thumbs up/down feedback to adjust sampling parameters
  • Privacy posture: No login, API key stays in localStorage, telemetry is opt-out, chat history never touches a server
  • AGPL-3.0 with explicit “no enshittification allowed” clause

Caveats

  • The “best response” selection in GODMODE CLASSIC and the 100-point composite scoring in ULTRAPLINIAN aren’t explained in detail; the README describes the mechanism but not the evaluation criteria
  • Chat history is localStorage-only with no sync; the README repeatedly warns you’ll lose everything if you clear browser data
  • The optional API server has an opt-in “Open Research Dataset” feature that publishes your chats to a public HuggingFace dataset — off by default, but the PII scrubbing is explicitly “NOT guaranteed to catch everything”

Verdict

Worth a look if you’re doing LLM red-teaming research, prompt engineering, or just want a no-install multi-model playground with an adversarial bent. Skip it if you want a polished daily-driver chat UI or if the “liberated AI” branding makes you wince.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.