Articles
Longer write-ups from the team behind Ezkielyna API: what we measured on our own gateway traffic, what broke and how we fixed it, and how to pay for AI APIs without a credit card. For step-by-step setup of a specific tool, see the setup guides.
"OpenAI-compatible" is a spectrum: 10 quirks we hit running a multi-upstream gateway
tool_calls: null, three reasoning dialects, usage chunks with or without choices, cache-token fields with different meanings, and seven more. What breaks, why, and the defensive fix for each.
· 7 min read
Measuring LLM latency honestly: TTFT, total time, and why tokens/s needs at least 50 tokens
How we compute time to first token, total latency, tokens per second and uptime for every model on a gateway from real traffic, the rules that keep the numbers honest, and a 40-line script to measure your own route.
· 6 min read
Nine MCP servers, zero dependencies: what we learned writing the protocol by hand
One .mjs file per server, no SDK, Node 18+, 123 tests against local mocks. Why we did it, the protocol details that bit us, and when you should just use the SDK instead.
· 6 min read
Paying for AI APIs without a credit card: what works, country by country
Credit card ownership is 3-6% of adults in Indonesia, the Philippines, Vietnam, India and Nigeria. Here is what the big AI API billing pages accept, and the legitimate options left when you have no card.
· 7 min read
What developers on one Indonesian AI gateway used last week (5-12 Nov 2026)
100,722 requests across 20 models: DeepSeek took 88% of requests, prompts outweighed outputs 158 to 1, and coding agents are visible in the shape of the traffic. One gateway's data, with its limits stated up front.
· 5 min read
Bahasa Indonesia: Artikel dalam Bahasa Indonesia