The Analytics Layer of AI

Carbon turns your AI usage into detailed, actionable insights — so engineering and product decisions stop running on vibes

Works withOpenAIOpenAI SDKAnthropicAnthropic SDKVercelVercel AI SDKClaudeClaude CodeCodexCodex
Carbon AppExplore the demo

Find answers
to every question

You shouldn’t have to play detective with your own AI usage. Carbon helps you find the answers to your questions in seconds, not weeks.

What changed after Tuesday's deploy?
Why did our AI bill double this month?
Which agent burned $400 last Tuesday?
Which tool calls keep failing?
Which model handles long context cheapest?
Which threads run the longest?
Who are our heaviest users, and what do they cost us?
Which model is actually cheapest for this task?
Are retries double-counting our usage?
Where do agents stall mid-run?
Which provider failed most this week?
What did that support thread cost?
Is the new prompt actually cheaper?
Which feature is driving token spend?
What's our cost per active user?

One platform for
every AI model

  • OpenAIOpenAI
  • AnthropicAnthropic
  • GoogleGoogle
  • VercelVercel
  • MistralMistral
  • OpenRouterOpenRouter
  • DeepSeekDeepSeek
  • AlibabaAlibaba
  • MinimaxMiniMax
  • Xiaomi

The complete picture

See the complete picture of your AI usage, find patterns, and make sense of it all through beautiful, detailed charts on the Carbon dashboard

Click any card to open that page in the live demo

Integrate in
two lines.

Wrap the AI SDK you already use — OpenAI, Anthropic, Vercel AI. Two lines, and your first event lands in the dashboard in seconds.

import OpenAI from "openai";
import { Carbon } from "@carbon-js/sdk";
import { wrapOpenAISdk } from "@carbon-js/sdk/ai";

// Wrap your client with Carbon SDK.
const carbon = new Carbon();
const openai = wrapOpenAISdk(new OpenAI(), carbon);

// Use exactly as before — every call is captured.
const completion = await openai.responses.create({
  model: "gpt-5.5",
  input: "Hello, Carbon!",
});
npx @carbon-js/connect@latest

Or in zero.

Running Claude Code or Codex? One command to wire up your AI apps to Carbon.

ClaudeClaude Code
CodexCodex

Start free.
Scale when you're ready.

Free forever. No credit card. Upgrade when your traffic does.

Carbon Free

$0/ month

For personal projects

Features

1,000 events per month

Additional events not supported

LLM and tool call tracking

All providers and SDKs supported

All clusters included

Carbon Plus

$28/ month

For production projects

Features

100,000 events per month included

$0.00028 per additional event

LLM and tool call tracking

All providers and SDKs supported

All clusters included

Questions?

Everything here describes how Carbon works today — not a roadmap.

How is Carbon different from tracing tools like Langfuse or Helicone?

Those are debuggers — built to inspect one call or one trace at a time. Carbon is the analytics layer above the whole population: cost, usage, latency, and adoption across every agent, model, and user, sliceable any way you ask. Plenty of teams run both — a tracer to debug a single request, Carbon to answer questions about all of them.

Will the SDK slow down my app?

No. Capture is non-blocking — events are validated, buffered, and batched off your request's hot path, then sent in the background. If Carbon is ever unreachable, your app keeps running as if it weren't there.

What happens to my prompts and outputs?

Everything is scoped to your private space — only you can query it. The full event, inputs and outputs included, is stored so you can inspect any call in full; per-field controls over what's retained are on the way.

Which providers and SDKs are supported?

Our SDK wraps the SDKs you already use — OpenAI, Anthropic, Vercel AI, and more — and coding agents like Claude Code and Codex connect without any SDK at all.

Does Carbon work with Claude Code and Codex?

Yes — run npx @carbon-js/connect and pick the apps to connect. It points their built-in OpenTelemetry export at Carbon, so usage, cost, and tool calls stream into your space. Restart the app and you're live.

How much code does it take to try?

Two lines with your existing AI provider — or none at all: a single command connects Claude Code or Codex. Either way your first event lands in the dashboard in under five minutes.

What counts as an event?

One captured LLM call or one tool call.

What happens if my included usage is reached?

On the Free plan, capture pauses until the next period starts — you can upgrade to Plus anytime to keep going. On Plus, events keep flowing and you're simply billed for the overage.

Can I take my data out of Carbon?

Yes, you will be able to export the events very soon. We are working on a way to implement this.

~~~ Let's begin ~~~

Two lines of code — or one command — a thousand free events a month, and every decision after that comes with evidence.