RepoDaily · 2026-07-20 · Infrastructure / Runtime

G0DM0D3: A Single-HTML Multi-Model Chat Built for AI Red Teaming and Cognition Research

#16 Infrastructure / Runtime TypeScript +272 elder-plinius/G0DM0D3 Open repository

elder-plinius's TypeScript chat interface races up to 60 OpenRouter models in parallel, runs on Ollama, and ships as one index.html file with no build step.

Repo typeInfrastructure / Runtime
Best forRed teamers, prompt engineers, and cognition researchers who want a no-install browser UI for parallel multi-model probing, input perturbation, and local-model racing.
Risk levelMedium — AGPL-3.0 copyleft, metadata telemetry on by default, and an anonymous maintainer mean review before any team deployment.
Time to evaluateUnder 10 minutes to open the hosted site or clone and open index.html; 30 minutes to wire Ollama and test ULTRAPLINIAN locally.

Primary question: Does your team need a transparent, self-hostable multi-model racing UI strong enough for adversarial prompt testing, and can you accept AGPL-3.0 obligations?

90/100

RepoDaily adoption score

RepoDaily rates this as 90/100 (strong) for adoption: evidence, installation path, production risk, differentiation, license clarity, and AI/agent fit are scored from the article sources and adoption notes.

Directional score from RepoDaily sources and adoption notes, not a benchmark.Risk: Medium
100Evidence quality

5 source(s) across 3 source category/categories, plus a RepoDaily-specific evidence module when available.

100Installability

6 workflow step(s), 5 next-action step(s), and 4 command/install signal(s) were detected.

61Maintenance confidence

Trending momentum is +272 stars, with maintenance/release/issue signals counted when present.

96Production readiness

Risk is marked medium, with 7 security note(s) and 4 explicit skip condition(s).

100Differentiation

3 opportunity lens item(s), 4 alternative(s), and 4 type-specific section(s) support differentiation.

82License clarity

License source or license wording is present.

72Agent / AI fit

4 AI/agent-related signal(s) were detected in the article text and metadata.

Project overview

G0DM0D3 is a fully open-source, privacy-transparent chat interface created by elder-plinius, the handle of a well-known AI red-teaming researcher. The README positions the tool for hackers, philosophers, and system tinkerers who want to push the post-training layer of language models. The hosted instance lives at godmod3.ai and accepts an OpenRouter key, a Venice key, or a local OpenAI-compatible server endpoint.

The project is unusual because its core UI is a single index.html file with no build step and no package install. Despite that, it exposes advanced features that mainstream chat clients omit: GODMODE CLASSIC races 5 prompt-plus-model combinations in parallel, ULTRAPLINIAN scales that race to between 12 and 60 OpenRouter models across 5 tiers with composite scoring, and Parseltongue applies 33 input perturbation techniques across 3 intensity tiers for adversarial testing.

The repository is TypeScript at the top level, but the CONTRIBUTING.md reveals the development stack is Next.js plus React in src/, an Express API proxy in api/server.ts, and a standalone Express app in HF/ for the Hugging Face Space build. The Dockerfile deploys a Research Preview API on node:20-slim, listening on port 7860, with OPENROUTER_API_KEY and GODMODE_API_KEY as required environment variables.

Because the maintainer is anonymous and the framing is adversarial, this is not a generic enterprise chat client. It is closer to a research instrument for probing model behavior, benchmarking prompts, and running local models through the same evaluation harness as commercial APIs.

Problem it solves

  • Most multi-model chat UIs hide their system prompts and sampling logic, making adversarial comparison hard to reproduce.
  • Running many models in parallel usually requires custom scripting or paid playgrounds with opaque scoring.
  • Red teamers need repeatable input perturbation, not just manual rephrasing, to stress-test guardrails.
  • Privacy-conscious users want a UI that does not silently exfiltrate conversations to a third party.
  • Self-hosting an evaluation harness often means wrestling with Docker, databases, and build pipelines before any testing begins.

How it works

  1. Visit godmod3.ai or clone the repo and open index.html directly in a browser — no npm install required for the static UI.
  2. In Settings, enter an OpenRouter API key, a Venice key, or a local OpenAI-compatible endpoint such as http://localhost:11434/v1 for Ollama.
  3. Pick a mode: standard chat, GODMODE CLASSIC for 5 parallel combos, or ULTRAPLINIAN for a 12-to-60-model race with composite scoring.
  4. Optionally enable Parseltongue to perturb your input with 33 techniques, and AutoTune to adjust sampling parameters across 20 query contexts.
  5. Conversations and settings persist in browser local storage with export and import support; Local-only mode disables OpenRouter, Venice, and app telemetry.
  6. To reproduce the hosted telemetry pipeline, deploy functions/api/telemetry.ts on Cloudflare Pages and configure its Hugging Face variables.

Architecture: Single HTML Front End, Express API, Cloudflare Telemetry

The production UI is one index.html file, which is the key architectural decision. That file calls OpenRouter, Venice, and local OpenAI-compatible servers directly from the browser. The README confirms a static-only deployment can run the chat UI, but /api/telemetry will not publish metadata unless functions/api/telemetry.ts is deployed on Cloudflare Pages.

CONTRIBUTING.md adds the development reality: the source lives in src/ as Next.js plus React plus TypeScript, the API proxy is an Express server in api/server.ts, and the Hugging Face Space build is a standalone Express app in HF/. The Dockerfile builds a Research Preview API from api/, src/lib/, and src/stm/, exposing port 7860 and running npx tsx api/server.ts.

This split means there are effectively two artifacts: a zero-dependency static front end for end users, and a containerized API for researchers who want the hosted evaluation pipeline, Hugging Face dataset publishing, and bearer-token authentication.

Command Surface: Clone, Serve, Pull, Deploy

  • git clone https://github.com/elder-plinius/G0DM0D3.git — clone the repository.
  • python3 -m http.server 8000 — serve the static UI locally after cd G0DM0D3.
  • ollama pull qwen3:8b and ollama serve — prepare and run a local model for Settings → API Keys → Local Models at http://localhost:11434/v1.
  • npm install and npm run dev — developer workflow described in CONTRIBUTING.md.
  • docker build -t g0dm0d3-api . and docker run -p 7860:7860 with OPENROUTER_API_KEY and GODMODE_API_KEY — build and run the Research Preview API container.
  • GET /v1/health — container health check endpoint used by the Dockerfile HEALTHCHECK directive.

Deployment Notes: Static Hosts and Cloudflare Pages

The README lists GitHub Pages, Vercel, Cloudflare Pages, Netlify, or any web server as valid static hosts for index.html. The only piece that requires a specific platform is the telemetry function, which must be deployed on Cloudflare Pages as functions/api/telemetry.ts.

The Dockerfile targets Hugging Face Spaces Docker SDK or any container host, expects port 7860, runs as a non-root user created with addgroup --system app and adduser --system --ingroup app app, and includes a 30-second health check interval. Required environment variables are OPENROUTER_API_KEY for model calls, GODMODE_API_KEY for bearer authentication, and optionally HF_TOKEN and HF_DATASET_REPO for auto-publishing data to a Hugging Face dataset repository.

Maintenance Risk: Anonymous Maintainer, Version 0.4.x, AGPL-3.0

SECURITY.md states that only version 0.4.x is supported; anything below 0.4 is explicitly unsupported. The vulnerability response timeline is 48 hours for acknowledgement, 7 days for initial assessment, and 30 days for a fix on critical issues, which is a reasonable but non-binding commitment.

The project is licensed under AGPL-3.0, confirmed by both the LICENSE file and CONTRIBUTING.md. Anyone operating a modified version on a publicly accessible server must release source code to users of that server, which affects commercial deployment planning.

The maintainer identity is the anonymous handle elder-plinius. The SECURITY.md scope references src/ and Docker configuration, but the README describes a single-file front end, so reviewers should verify which artifact actually ships before relying on the security controls.

Who should pay attention?

Good fit if

  • Red teamers who need repeatable perturbation across 33 techniques without writing custom scripts.
  • Researchers who want to race 12 to 60 OpenRouter models and compare composite scores.
  • Privacy-focused users who want to run everything against local Ollama or llama.cpp models.
  • Developers who value a no-build single-file front end they can drop on any static host.

Skip for now if

  • Teams that cannot accept AGPL-3.0 network copyleft on a modified public deployment.
  • Enterprises that require a verified legal entity and vendor SLA behind a chat tool.
  • Users who need turnkey SSO, audit logs, and admin controls out of the box.
  • Anyone uncomfortable with metadata telemetry enabled by default on the hosted site.

Risks and cautions

Medium

The codebase is transparent and self-hostable, but AGPL-3.0, default telemetry, and an anonymous maintainer require due diligence before team adoption.

  • AGPL-3.0 imposes source disclosure obligations on modified versions served over a network.
  • Metadata-only app telemetry is on by default and must be explicitly disabled via No-Log or Local-only mode.
  • The maintainer is the anonymous handle elder-plinius, with no named organization behind the project.
  • SECURITY.md only supports version 0.4.x, so older forks or cached deployments may be unsupported.
  • The scope mismatch between SECURITY.md referencing src/ and the README describing index.html means reviewers must confirm which artifact they are actually deploying.
  • Authentication uses a bearer token with constant-time comparison to resist timing attacks, per SECURITY.md.
  • Rate limiting is tier-aware with per-minute, per-day, and lifetime sliding windows.
  • Security headers include HSTS, CSP, X-Content-Type-Options, X-Frame-Options, and Permissions-Policy.
  • Docker containers run as a non-root user on minimal base images with a health check on /v1/health.
  • The data policy states zero PII storage and opt-in dataset collection only.
  • Conversations and settings persist in browser local storage, not on a remote server, unless telemetry is explicitly configured.
  • Vulnerability disclosure is via private report only; SECURITY.md asks users not to open public issues for security bugs.

Alternatives to compare

ApproachWhen to useTrade-off
LibreChat
You want a more conventional multi-provider chat UI with user accounts, plugins, and a broader contributor base.MIT licensed, free and open source.
Open WebUI
Your priority is a polished Ollama-first front end with document tools and RBAC rather than adversarial red-teaming features.MIT licensed, free and open source.
AnythingLLM
You need a RAG-focused workspace with document chat and enterprise workspace isolation.MIT licensed, free and open source.
Chatbox
You want a lightweight cross-platform desktop client for multiple providers without parallel model racing.Free and open source.

What this trend reveals

Internal Red-Teaming Workbench

Security teams can self-host G0DM0D3 on an internal static server, point it at local Ollama or vLLM instances, and use Parseltongue's 33 perturbation techniques as a repeatable guardrail test suite.

Run 5 representative prompts through Parseltongue at intensity tier 3 against your guardrailed internal model and check whether any variant bypasses policy.

Model Procurement Benchmark

Procurement teams evaluating multiple OpenRouter models can use ULTRAPLINIAN's composite scoring across the 12-model tier to compare candidates on the same prompt set.

Pick 10 business-critical prompts, run them through ULTRAPLINIAN tier 1, and export results from browser local storage for side-by-side scoring.

Privacy-First Local Assistant

Users who want zero external calls can enable Local-only mode, which excludes OpenRouter and Venice and automatically disables G0DM0D3 app telemetry, leaving all traffic between browser and localhost.

Confirm with browser dev tools that no outbound request leaves your machine after enabling Local-only mode and pointing Settings at http://localhost:11434/v1.

Best next action

Self-Host the Static UI and Point It at a Local Ollama Model

The fastest way to evaluate G0DM0D3 without sending data anywhere is to clone the repo, serve index.html locally, connect Ollama, and run one ULTRAPLINIAN race.

  1. Run git clone https://github.com/elder-plinius/G0DM0D3.git and cd G0DM0D3.
  2. Start a local server with python3 -m http.server 8000 and open the URL in your browser.
  3. In a second terminal run ollama pull qwen3:8b then ollama serve.
  4. In Settings → API Keys → Local Models, enter http://localhost:11434/v1, click Test & Discover Models, and enable Local-only mode.
  5. Add at least 3 model IDs and launch an ULTRAPLINIAN race on a test prompt to confirm the pipeline works end to end.

RepoDaily verdict

G0DM0D3 is a genuinely distinctive open-source chat client aimed at red teamers and cognition researchers, not a general-purpose assistant UI. Its single-file deployment, parallel model racing, and 33-technique perturbation engine justify the attention, but AGPL-3.0, default telemetry, and an anonymous maintainer mean it belongs in a research or internal-evaluation context before any broader rollout.

Sources