
The story behind discode.ai
A coffee house co-owner
who can’t code
built discode.ai
Well, to be precise: 49.6% of his code is still in there ;-) More than half of the vibe-code was refactored and rebuilt by the pros on the team. But let’s start at the very beginning.
Since Large Language Models conquered the internet, Moriz threw himself in with enthusiasm. What started two years ago with ChatGPT quickly turned into cross-checking chaos across platforms. ChatGPT as the incumbent, Gemini for Google Workspace, the new darling Claude for its large context window. Result: three subscriptions in parallel — first €64 a month, then hundreds. Redact personal data across thousands of PDF pages. And second thoughts about sustainability.
Copy. Paste. Switch tabs. Repeat.
Fall 2025. New models with new features hitting the market faster and faster — at some point too much to keep track of in your head. And a nagging thought: what does all this computing power actually cost the planet? A new solution was needed.
He built himself a personalized GPT loaded with weekly updated benchmarks that recommended the best model per task. It worked — but still too much back-and-forth in everyday prompts.
Mo wanted to automate. He spent four to five days over the Christmas holidays on a no-code platform and left disappointed — still not what he was looking for. So he sent screenshots to an old school friend and coding wizard working at Google in San Francisco: “What kind of developer do I need?”
The former classmate: “None. You don’t need a developer anymore. We’ve all been playing with Claude Code. Try it!”
Mo was skeptical — but if his old Monkey Island play-partner at Google says so, it can’t hurt to try. After two days, he had better results than anything before. Pandora’s box was opened.
“Use Claude Code to build
something more responsible than Claude?
If that’s not the plot twist of the AI age,
what is?”
Moriz, Initiator of discode.ai
Without a coding background — the last touchpoint being an HTML continuing education course in the late nineties — and with modest math from school, Mo put in the work and kept refining his idea. Not in variables. In sentences. In images.
What followed were weeks of double life. Running the coffee house by day, vibecoder by night. Always deeper into the rabbit hole — three Claude Max subscriptions and $2,000 in API tokens later. (We’re hoping your usage helps us offset that carbon footprint.)
With a little help from his friends.
Despite the love for his new pastime, it soon became clear: you can’t ship without pros. Over several weeks, an illustrious crew of frontend, backend, product design, and graphic design professionals came together — and sculpted Mo’s wild 417,000-line monster into a lean 185,000-line discode. To be precise: 49.6% of his original code is still in there.
A coffee house co-owner, an idea, and an AI that speaks code. 100+ models. Automatic routing. The world’s first AI Eco Score. A disco ball that shines light into a dark room — making the AIs dance to your tune.
You choose the rhythm, not the algorithm.
Have fun! Tell us what you think! — Vroni, Sebastian, Harry, Alexei, Michi, Peter & Moriz




GitHub Contributions Heatmap
Why discode?
The big AI platforms from the US and China build for themselves. Who builds for Europe? No platform combines consumer UI with cost transparency. No provider shows a visible Eco Score.
Intelligent Orchestration
Smart routing — the right model for every request, automatically. 60–70 % cost almost nothing.
100+ Models — 1 Chat.
discode chooses the best AI model for you. Finally clarity in the engine madness. Goodbye AI-Fomo.
Eco Score — World First
CO₂, water, electricity — visible per request. No other provider shows you what AI really consumes. AI energy consumption will be the mainstream topic of 2025/26.
Anonymization Assistant
Data is anonymized locally before it leaves your device.
Stack
- •Next.js 16 (App Router) + React 19 + TypeScript strict
- •Supabase (PostgreSQL + RLS, 13 Tables, 32 Policies)
- •Clerk Auth (JWT, Server-side Validation)
- •Stripe (Checkout, Webhooks, Subscription Lifecycle)
- •Vercel Pro (EU Region, Edge + Serverless)
- •Redis/Upstash (Rate Limiting, Caching)
- •Vercel AI SDK 6.0 (SSE Streaming)
- •Tailwind 4 + shadcn/ui + Zustand 5
- •Playwright (343 Unit + 248 E2E Tests)
- •GitHub Actions CI/CD