Hacker News
Orbit is a new analytics tool designed to help AI teams monitor and manage the costs and performance of their AI agents in production, specifically by identifying "zombie loops" and calculating cost-per-feature.
AI 生成摘要
Orbit 是一款新推出的分析工具，旨在協助 AI 團隊監控和管理其 AI 代理在生產環境中的成本與效能，特別是透過識別「殭屍迴圈」並計算每項功能的成本。
In production, AI doesn't fail loudly. It fails expensively — through silent loops, prompt changes, and features that look fine but quietly burn margin.
Usage spike detected: $2,847 in the last 24h (+312%)
Why is our AI bill 3x higher this month? We need answers before the board meeting.
Checked the logs. We made 847k API calls yesterday, mostly to gpt-4o. But the logs don't tie back to features — I'd have to dig through code paths manually.
Was it the new summarization feature? The chat widget? A looping workflow? I genuinely don't know.
Orbit ties every call to a task. Every task to a customer.
So you always have the answer.
Here's how we solve it
Vendor dashboards show totals. Logs show requests. Neither tells you which feature is burning margin or why.
Orbit ties every LLM call to a product feature — and soon, to entire workflows — so you can see what is expensive, slow, or failing in production.
AI spend rarely grows linearly. One prompt tweak or model swap can turn a healthy feature into a loss overnight — and as systems become more workflow- and agent-driven, cost can explode in minutes.
Orbit tracks traffic vs. spend in real time and warns you the moment your unit economics break.
See error rates by feature and model. Understand which parts of your product are breaking and why — with detailed error logs and failure reasons.
Track error trends over time. Catch regressions early. Know exactly which model and feature combination is causing issues.
Orbit shows you what's actually happening in your application — without proxies, scraping, or hidden assumptions.
Metrics are captured directly from your application runtime. No external monitoring or traffic interception.
Orbit never sits between your app and your AI provider. Your requests go directly to OpenAI, Anthropic, etc.
Cost, latency, and error rates are calculated from real request data — not estimates or statistical sampling.
Your provider API keys stay in your application. Orbit only receives usage metadata, never credentials.
Provider dashboards show total API calls. They don't show which feature failed, why, or what it cost you.
One npm package. Wrap your AI client. See your data instantly.
Supported: OpenAI, Anthropic, Gemini|Coming soon: Mistral, Groq, DeepSeek
npm install @with-orbit/sdk
One line to instrument OpenAI
Real-time metrics in your dashboard
See what your AI is really costing you in production — feature by feature, workflow by workflow.
Feature-level AI observability using real runtime data.
© 2026 Orbit. All rights reserved.