Prompt Inspector

Before you spend money asking GPT-5.6…let Kimi inspect your prompt first.

Prompt Inspector automatically picks the cheapest GPT-5.6 tier that can nail your prompt — Luna for the small stuff, Terra for the workhorse jobs, Sol only when it's actually earned.

Try an example:

Point your SDK at it

One-line migration. Your code already speaks OpenAI — the proxy speaks it back, just cheaper.

01
Start the proxypnpm dev in apps/proxy — it listens on localhost:4000 and speaks the OpenAI API.
02
Change the baseURLThe only line of your code that changes. Every SDK that speaks OpenAI already works.
03
Ask for gpt-5.6-autoThe router inspects each request and picks the tier. Pin gpt-5.6-sol whenever you disagree.
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://localhost:4000/v1", // ← the only change
});

const res = await client.chat.completions.create({
  model: "gpt-5.6-auto", // Kimi picks the tier
  messages,
});

Everything the bill was hiding

Inspection, routing, and accounting — in one open-source box.

Inspect a Prompt

Paste any prompt, get a tier verdict with the reasoning to back it up — before you spend a cent.

Cost Comparison

Luna, Terra and Sol priced side by side on every request, down to the tenth of a cent.

Prompt Optimizer

Mechanical rewrites that drop your prompt a tier — savings verified by re-inspection, not vibes.

API Auto Router

An OpenAI-compatible proxy. Change one baseURL line and every request rides the cheapest sufficient tier.

Spending Dashboard

Every routed request logged with cost, savings and latency. Watch the Sol money come back.

Decision Inspector

Each routing decision ships with human-readable reasons. No black box, no hand-waving.

Challenge the Router

Run one prompt against all three tiers side by side and judge whether the cheap model keeps up.

Waste Detection

Flags traffic riding Sol that the classifier says belongs downstairs — priced per month, in dollars.

Benchmark Runner

30 labeled prompts across 10 categories. Prove the router agrees with humans before you trust it.

Built by a very large Chinese model with strong opinions about your OpenAI bill. Open source, MIT.