MecutinAI is the AI API gateway for markets global providers ignore. 18 models (DeepSeek, Qwen, Kimi, GLM) via one OpenAI-compatible endpoint. No VPN, no USD card — pay with QRIS (Indonesian e-wallets), API key active in 30 seconds. From Rp 5.000 (~$0.30). Live: 234 orders, 124 buyers, 98K+ weekly requests. Plus Agency white-label program and Showcase gallery. Built by an Indonesian developer who was tired of VPN tunnels dropping mid-request.
What inspired you to build this?
As a developer in Indonesia, I kept hitting the same wall: every time I wanted to build something with AI APIs, I'd hit a geo-block. OpenAI doesn't officially support Indonesia. Anthropic? Same story. The workaround was always the same painful loop — VPN that drops mid-request, prepaid USD credit cards that get declined, and crypto payment services that add 3 layers of friction.
I remember one night, my production chatbot went down because the VPN tunnel timed out during a streaming response. That was the moment I thought: why am I tunneling through three countries just to call an API?
I had access to AI provider partnerships (Alibaba Cloud/DashScope, etc.) with legitimate volume agreements. The infrastructure was already there. What was missing was a local layer — an Indonesian-facing gateway that speaks the same OpenAI-compatible protocol developers already know, but with QRIS payment, rupiah pricing, and zero-VPN access.
That's how MecutinAI was born.
What problem were you trying to solve?
Three core problems:
1. Access friction. Indonesian developers couldn't use AI APIs without VPN, USD credit cards, or crypto. I wanted to make it as simple as buying a GoPay top-up — scan QRIS, get API key, start building.
2. Cost barrier. OpenAI charges $5-20 minimum just to start, plus foreign transaction fees. For a student or indie developer in Indonesia earning in rupiah, that's a significant commitment before even testing an idea. I wanted entry-level pricing at Rp 5.000 (~$0.30) — enough to build a real prototype.
3. Fragmentation. Developers who wanted to use multiple models (DeepSeek for coding, Qwen for multilingual, Kimi for reasoning) had to manage multiple API keys, multiple billing accounts, multiple SDK integrations. I wanted one endpoint, one key, 18 models — switch models by changing one parameter.
How did your approach or process evolve while working on this launch?
The first version was embarrassingly simple — a single proxy file that forwarded requests to one provider. No auth, no quota, no billing. Just "proof it works from Indonesia without VPN."
Once I confirmed the core worked, the evolution happened in phases driven by real user feedback:
Phase 1 — Gateway & Quota: Built the real gateway engine with API key auth (bcrypt), dual-layer quota system (5-hour sprint window + weekly cap to prevent abuse), weighted token counting (premium models cost more), and SSE streaming with keepalive. The dual-layer quota was actually born from an abuse incident — users were cycling sprint resets 24/7 to get effectively unlimited access. The weekly cap was the response.
Phase 2 — Payment & Trust: QRIS payment integration with AI-verified proof approval. This was the hardest part — Indonesian payment proof is a screenshot, not a webhook. I built a multi-model vision AI verification pipeline (qwen + kimi consensus) with image forensics to detect edited screenshots. Auto-approve if confidence is high, manual review fallback otherwise. This cut order-to-key delivery from hours to ~30 seconds.
Phase 3 — Ecosystem: Added referral program, Telegram bot as a full payment channel (not just notifications), free trial via Telegram group join, key recovery via email OTP, and the Agency white-label program (source code + master key for resellers).
Phase 4 — Community & SEO: The Showcase feature ("Built with MecutinAI") lets users show off projects — community upvotes, unique click tracking, featured slots. Then 10 SEO landing pages targeting Indonesian search keywords ("api deepseek murah", "alternatif openai api", "api ai tanpa vpn") with FAQ structured data for Google rich snippets.
The biggest lesson: start with the friction, not the feature. Every phase was driven by a specific pain point — geo-block, payment friction, quota abuse, churn, discoverability. Features that didn't solve a real friction never shipped.