Loading your trial...
30-day AI usage & cost analysis
Add one small observer around your existing model client. Your private report starts updating immediately. Keep it running for 30 days to build a reliable cost baseline and uncover opportunities in model choice, caching, context size, and reasoning use.
No account required · No credit card · Metadata only by default
Your prompts, responses, and provider API keys stay in your app.
Two small changes
Keep your existing provider, prompts, and model calls. Wrap the client once and the observer records metadata in the background.
Remove the wrapper at any time to roll back.
import OpenAI from "openai";
const openai = new OpenAI();
// Your existing callsimport OpenAI from "openai";
import { observeOpenAI } from "@usagetap/observer";
const openai = observeOpenAI(new OpenAI());
// Your existing callsThe same wrapper supports OpenAI and OpenRouter's OpenAI-compatible client setup.
Value from the first call
The private report starts with connection status, calls, tokens, estimated spend, model mix, latency, and cache behavior. Feature labels add cost-by-workload detail. Thirty days makes the patterns more reliable; you do not wait 30 days to see them.
Projected savings are a cost estimate, not a quality claim. Validate with an opt-in representative sample before changing production.
Metadata only by default
UsageTap records provider and model; input, output, cached-input, cache-write, and reasoning token counts when the provider exposes them; response time; success or error status; timestamp; and an optional feature label.
We do not collect prompts, model responses, provider API keys, messages, or tool content.
Prompt content is never sent for compression automatically. If you explicitly choose a prompt for a compression test, only that selected prompt is sent. Your production model call is never altered by the trial. You choose every test, with a 50,000-token per-request limit and a one-million-input-token trial budget.
What you get
Metadata can show which models, token categories, and workloads drive cost. It can surface cheaper-model and caching candidates without changing production behavior. Content quality and unnecessary context require a separate, explicit sample test.
See every result without creating an account.
Enter your email once for setup or later for a fresh private report link.
If the trial proves useful
The lightweight observer does not identify individual users. Customer and user-level attribution begins only when you deliberately move to the full SDK and provide your own stable identifiers. The live trial report itself never requires an account.
Let the live report update, add stable feature labels, and optionally test bounded compression on a suitable workload.
See connection status, usage, model mix, response times, cost, and candidate savings through a private email link. No account required.
Create an account only to keep or share the workspace, then move to the full SDK for customer and user IDs, limits, alerts, forecasts, compression, or routing.