elder-plinius/G0DM0D3

What it does

G0DM0D3 is a browser-based chat interface that connects to OpenRouter, Venice, or any OpenAI-compatible local model server. It is aimed at people who study how models respond to adversarial prompts: red teamers, researchers, and system tinkerers. The standalone app is a single index.html served as a static file; the repository also contains an optional React/Next.js frontend and a Node/Express API server.

Two multi-model engines are central. GODMODE CLASSIC runs five fixed model-plus-jailbreak-prompt combinations in parallel and keeps the best response. ULTRAPLINIAN queries larger model pools in tiers (12 to 60 OpenRouter models, plus Venice and local models when configured) and scores each response on a 100-point composite metric. A Parseltongue module applies 33 obfuscation techniques to inputs at three intensity levels, and AutoTune picks sampling parameters based on one of 20 query contexts.

Get started

The hosted version is at godmod3.ai; you supply an OpenRouter key, Venice key, or a local endpoint. To self-host the standalone UI:

git clone https://github.com/elder-plinius/G0DM0D3.git
cd G0DM0D3
python3 -m http.server 8000

Open the local URL in a browser, then configure providers under Settings. A purely static deployment works for chat, but the /api/telemetry endpoint only functions if you also deploy functions/api/telemetry.ts on Cloudflare Pages.

To use local models with Ollama:

ollama pull qwen3:8b
ollama serve

Then go to Settings → API Keys → Local Models, enter http://localhost:11434/v1, and click Test & Discover Models. Additional model IDs can be added to race together in ULTRAPLINIAN. If you use the hosted site with a local server, that server must allow browser CORS from https://godmod3.ai. Setup details for Ollama, LM Studio, and vLLM are in LOCAL_MODELS.md.

Data and privacy constraints

The app has no accounts. API keys and chat history are stored in browser localStorage; clearing browser data or switching devices loses conversations unless you export them first. Hosted app telemetry is metadata-only and on by default, reporting session IDs, timings, model/mode/tier, scores, content lengths, and classification labels to a Cloudflare Pages function. When telemetry is enabled, the raw prompt is also sent to an auxiliary model to generate a harm-taxonomy label, with only the label included in the telemetry event. No-Log Mode disables app telemetry; Local-only mode additionally excludes OpenRouter and Venice calls.

The optional API server has a separate opt-in full-content dataset path triggered by contribute_to_dataset: true. There is no automatic PII scrubber on that path, and contributed messages may be published to a public Hugging Face dataset. The standalone hosted app does not expose this flag. The full policy is in TERMS.md.

Limits and license

The project is explicitly a red-teaming tool, and its jailbreak combinations target specific model versions that may change or be unavailable over time. The hosted single-file UI and the optional React frontend differ in features; the README warns not to assume features from src/ exist in the hosted index.html. The code is licensed AGPL-3.0, so modified versions offered over a network generally must make their source available. API and endpoint details are in API.md, and the design and evaluation write-up is in PAPER.md.

HotRepo's take

The project appeared on the daily all-language list on 2026-07-18 with 63 new stars and 9,408 total, then 69 new stars the next day. By a 2026-11-12 TypeScript daily snapshot it had 12,221 stars and ranked 9th on that language list, with 27 stars gained that day. That is a small but persistent presence across two separated observation periods rather than a single spike. The repository does not have a recorded best rank on the all-language GitHub daily list, so its actual peak there is unknown; the 9th-place figure applies only to the TypeScript list on that date.