// the find
elder-plinius/G0DM0D3
LIBERATED AI CHAT
G0DM0D3 is a single-file, no-build chat UI for racing prompts across dozens of LLM providers at once (OpenRouter, Venice, local Ollama/LM Studio/vLLM), built around jailbreak-style prompt combos and red-teaming tooling. It's for people probing model safety boundaries and comparing raw model outputs side by side, not general end users wanting a daily chat client.
Self-hosting is genuinely zero-friction — index.html with no npm install or build step, which is rare for anything this feature-dense. Local model support is real, not a stub: OpenAI-compatible endpoints for Ollama/LM Studio/vLLM with CORS and discovery documented. The privacy section is unusually specific, listing exact telemetry fields rather than a blanket 'we don't track anything' claim. Racing up to 60 models in parallel with composite scoring (ULTRAPLINIAN) is a legitimate model-eval harness once you strip away the jailbreak framing.
The headline features — GODMODE combos and Parseltongue obfuscation — are jailbreak prompts hardcoded against specific model versions; providers patch these regularly, and the README itself admits listed models can go stale before the repo catches up. There are three overlapping, partially-diverging implementations (standalone HTML, Next.js frontend, Express API) with explicitly documented feature gaps between them and no visible test suite for any of them, so it's unclear which is canonical or how regressions would be caught. Telemetry sends raw prompt text to an auxiliary LLM for harm-taxonomy classification by default, which is a bigger privacy footprint than the 'metadata-only' framing implies unless you manually enable No-Log mode. The API server's opt-in full-content dataset contribution has no PII scrubbing beyond stripping system messages, so a careless integrator flipping that flag is a real leak vector.