Cosen - LLM-agnostic cost, traces, and security. Self-host it.

Vivek here, maker, from Pune. What pushed me to build Cosen was not a blank-canvas idea. It was the same week, over and over, while shipping AI features. I would pick a model on Monday, swap it on Wednesday when the bill jumped or the output broke, and by Friday I still could not answer three questions that a product person should be able to answer in one place: - What did *this feature* cost today? - Which call was slow or wrong? - Did we just send a key or a prompt-injection through? The stack for that was never one tool. Cost lived in a gateway. Traces lived in an observability product. Security was a scan I ran later, or not at all. Every model change meant the story fell apart again. I have spent years on product development, LLM architecture, application work, cloud and DevOps. The part that still felt unfinished was the control plane around the model — the layer that should not care whether you are on OpenAI, Groq, Ollama, or vLLM. So I started building public AI products for people who actually ship. Cosen is the first: one local web app and gateway. Tag a call with a feature name. You get spend, a trace, and a block or redact if the prompt looks like a secret or an injection. Mock mode needs no API key. I would rather you try `cos serve --mock`, open localhost:8080, and tell me what is missing than collect polite upvotes. If you are wiring AI into a real product, I want to hear which of those three questions still hurts. That is the list I will build next.