I built Token China — an API gateway for DeepSeek and GLM — as a solo developer. Here is the full tech breakdown.
Architecture
Nginx (HTTPS) → Go binary (embedded React SPA) → SQLite
- Backend: Go 1.26, forked from new-api (open source AI gateway)
- Frontend: React + Rsbuild + TanStack Router, embedded via Go embed.FS
- Database: SQLite (zero external dependencies)
- Hosting: Single Vultr VPS, Singapore ($6/month)
- Process: systemd, no Docker
Models
Four upstream channels through one OpenAI-compatible endpoint:
| Model | Category | Multiplier |
|---|---|---|
| deepseek-v4-pro | Reasoning | 1.75x |
| deepseek-v4-flash | Fast | 0.1x |
| glm-5.1 | General | 0.5x |
| glm-5v-turbo | Vision | 0.72x |
Payment
USDT (TRC20) deposits with automatic on-chain monitoring. A Python daemon polls TronGrid every 30 seconds and credits user accounts in real-time.
Lessons Learned
- Go embed.FS does not work with http.FileServer for SPAs — had to write a custom filesystem handler
- Docker OOM on a $6 VPS was a dealbreaker — native systemd reduced memory from 512MB to ~18MB
- DNS hijacking by ISPs is a real problem for .cc domains
Try It
token-china.cc — free starter credits on signup.
Top comments (0)