Get started
Introduction
What krino is, who it is for, and how it makes the small decisions your AI agent takes cheaper without changing what your agent does.
Your agent makes many small decisions on every run. Which tools does the model need for this task? Is this tool call safe to run? Today the big model decides by default, and you pay for it in tokens.
krino puts a small, cheap decision model next to your agent. It asks that model the same small questions, records the answers, and tells you what you would save by trusting them.
What krino does
- Watches. In shadow mode, krino asks the decision model in the background and records its answers. Your agent behaves exactly as before and waits for nothing.
- Prices.
npx krino reportreads the traces and shows, per decision, how often krino agreed with your agent and what it would save. Costs include prompt-cache reads and writes. - Cheapens, safely. When the data agrees, turn on enforce mode for tool selection. krino then sends only the tools the task needs. If the decision model is slow or unsure, your agent gets all its tools.
The two decisions in v0.1
| Decision | The question | When unsure |
|---|---|---|
| Tool selection | Which of the agent's tools does this task need? | Send all tools (fails open) |
| Risk gate | Is this tool call safe to run without asking a person? | Suggest asking a human (fails closed) |
The risk gate runs in shadow mode only in v0.1: it records what it would do and never blocks a call.
Who it is for
krino is for TypeScript agents built on the Vercel AI SDK or the Claude Agent SDK, especially agents with many tools, where tool definitions are a large part of every prompt.
What krino is not
- Not a proxy. krino is a library that runs inside your process. Nothing sits between you and your model provider.
- Not a model router. Your agent keeps its model. krino only answers small questions about tools.
- Not a hosted service. Traces are files on your machine.
Read next
- Install krino and the CLI.
- Follow a quick start: Vercel AI SDK or Claude Agent SDK.
- Learn the safe path to enforce mode: Shadow, report, enforce.