OpClave
Four minds. One answer.
OpClave sends your question simultaneously to GPT, Claude, Gemini, and Grok. A fifth model reads all four responses and synthesizes them into a single unified answer — surfacing where the models agree, flagging where they diverge, and producing a conclusion no single model could reach alone. That's Voyance, the core chat experience.
Studio is the companion workspace for heavier tasks:
- Build — generate a working single-file app or component, then iterate on it conversationally
- Research — a multi-model research explorer with live web sources, verified academic citations, and auto-generated charts
- Review — three specialized agents (security / quality / performance) audit a code snippet in parallel
Live at opclave.com
How it works
- You ask a question (or attach a file)
- All four models answer independently and in parallel
- A synthesis model reads every response and produces one unified answer with consensus, disagreement, and a final judgment
Two modes are available:
| Mode | Models | Use case |
|---|
| General | GPT-mini, Claude Sonnet, Gemini Flash, Grok Fast | Everyday questions, research, decisions |
| Intensive | GPT, Claude Opus, Gemini Pro | Coding, complex reasoning, technical tasks |
Stack
- Backend — FastAPI + Mangum (deployed as a Vercel serverless function)
- Frontend — Vanilla HTML/CSS/JS with Tailwind
- AI providers — OpenAI, Anthropic, Google Gemini, xAI Grok, Perplexity (web search + research)
- Auth & sync — Supabase (email/password auth, Postgres with row-level security for threads/artifacts)
- Rate limiting / logging — Upstash Redis
- Deployment — Vercel
Project structure
api/index.py # FastAPI app — all routes and model calls
templates/
index.html # Main app UI (Voyance + Studio)
landing.html # Marketing landing page
studio.html # Studio mode picker
preview_matrix.html # Consensus matrix preview/demo page
privacy.html # Privacy policy
assets/ # Fonts, logos, compiled Tailwind CSS
prompts/
synthesis_text.txt # Synthesis model system prompt (text)
synthesis_image.txt # Synthesis model system prompt (images)
public/ # Static landing assets served directly
src/tailwind.css # Tailwind source (compiled to templates/assets/tailwind.css)
supabase_schema.sql # Postgres schema + RLS policies for threads/artifacts
logs/queries.jsonl # Local query log fallback (gitignored)
Running locally
Prerequisites: Python 3.9+, Node (only needed if you're editing Tailwind styles), and an API key for at least one model provider.
git clone https://github.com/aarush-bajaj/OpClave.git
cd OpClave
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env # then fill in the values you have
See .env.example for the full list of keys. Only the model provider keys (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, XAI_API_KEY, PERPLEXITY_API_KEY) are required — Redis rate limiting, Supabase auth/sync, feedback email, and admin protection are all optional and degrade gracefully when unset.
If you want cloud-synced threads/artifacts and sign-in, run supabase_schema.sql in your Supabase project's SQL editor (Database → SQL Editor → New query) before setting SUPABASE_URL / SUPABASE_ANON_KEY.
uvicorn api.index:app --reload
Open http://localhost:8000.
Deploying to Vercel
The repo is structured for zero-config Vercel deployment. vercel.json rewrites all routes to api/index.py via Mangum.
vercel deploy
Set the env vars from .env.example in your Vercel project settings. In production, always set ADMIN_CODE — without it, the /logs admin endpoints are open to anyone.
Rate limits
| Bucket | Limit |
|---|
| General queries | 15 per IP (or account) per day |
| Intensive queries | 3 per IP (or account) per day |
| Feedback submissions | 3 per IP per hour |
| File uploads | 10 per IP per hour |
Limits are enforced via Upstash Redis. If Redis is unavailable, requests fail open.
License
MIT — see LICENSE.
Built by
Aarush Bajaj and Aditya Mishra
