See where your AI spend goes.

TypeScript SDK · Python SDK · HTTP API

See where your AI spend goes.

Marginal tracks the cost of every LLM call and slices it by the fields you care about — customer, feature, model.

Start trackingRead the docs

npm install marginal-sdk · pip install marginal-sdk

45.6% of spend

is one customer — found by grouping on a field you define

Priced server-side

provider + model + usage in, dollars out — no token math in your app

marginalhq.com

ExplorerLast 14 daysGroup by customerfeature = chat

Total spend

$1,686.24

Events

38,912

Top customer

acme-corp

$200$150$100$50$0

Aug 1Aug 5Aug 10Aug 14

acme-corpglobexinitechOther

| customer | Events | Spend | Share | | ---------------- | ------ | ------- | ----- | | acme-corp | 17,882 | $768.10 | 45.6% | | globex | 11,204 | $485.55 | 28.8% | | initech | 6,391 | $278.66 | 16.5% | | stark-industries | 2,145 | $96.41 | 5.7% | | hooli | 1,290 | $57.52 | 3.4% |

The dashboard

From “the bill doubled” to a named cause

Slice spend by what matters in your product, see every dimension charted automatically, and check exactly what your integration sent.

model

7d

feature

7d

Spot the spike without building a chart

Insights charts spend for every dimension automatically — each model, provider, and registered field gets its own chart. No setup, no query builder.

success500 accepted12:04:31

warn500 accepted · key “team” ignored12:04:08

error499 accepted · 1 rejected12:03:42

Know exactly what landed

Every API request is logged with its outcome — accepted counts, rejected events and why, stripped keys, unpriced models. Debug your integration without guessing.

gpt-4o-mini$0.15 / $0.60 per 1M

claude-sonnet-4$3.00 / $15.00 per 1M

internal-router-v2unpricedSet price

Never a silent $0

Unknown model? The event still lands, flagged as unpriced — visible in the dashboard and the API response, never averaged away. Set a custom price and it applies at ingest.

How it works

One call per LLM request. That's the whole integration.

1. 1

Pick your connection

Track each LLM call

Slice your spend

TypeScriptPythonHTTP

import { Marginal } from "marginal-sdk";
const marginal = new Marginal({ apiKey: process.env.MARGINAL_API_KEY });
const response = await openai.chat.completions.create({ / … / });
marginal.track({
  provider: "openai",
  model: response.model,
  usage: response.usage,
  fields: { customer: "acme-corp", feature: "support-bot" },
});

import os
from marginal import Marginal
marginal = Marginal(api_key=os.environ["MARGINAL_API_KEY"])
response = client.chat.completions.create(...)
marginal.track(
    provider="openai",
    model=response.model,
    usage=response.usage.model_dump(),
    fields={"customer": "acme-corp", "feature": "support-bot"},
)

curl -X POST https://api.marginalhq.com/v1/events \
  -H "Authorization: Bearer $MARGINAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "events": [
      { "provider": "openai",
        "model": "gpt-4o-2024-08-06",
        "usage": { "prompt_tokens": 2006, "completion_tokens": 300 },
        "fields": { "customer": "acme-corp", "feature": "support-bot" } }
    ]
  }'

The cost lands computed, priced at that day's catalog rates — no token math in your codebase.

Cost observability

Built to answer “why was the bill that high?”

Slice by any field

Register your vocabulary — customer, feature, anything you slice by — then group or filter spend by it. Unregistered keys are stripped and reported back, so dashboards stay clean.

LLM-aware pricing

Send provider, model, and the response's usage object; cost is computed server-side from a daily-synced price catalog, with per-project overrides. Prices are frozen at ingest.

Saved views

Any Explorer state — range, filters, group-by — saves as a named view. The questions you ask every week are one click away.

SDKs that stay out of the way

TypeScript and Python, zero dependencies. Events buffer locally and flush in batches in the background; network failures retry and then warn — your app never notices. Prefer no SDK? The HTTP API is a single JSON POST.

Let your coding agent do the integration

Paste marginalhq.com/llms.txt into Claude Code, Cursor, or any coding assistant and it has everything it needs to wire up Marginal — or follow the per-provider recipes yourself.

Integration recipes

From zero to a live spend dashboard in five minutes.

Create a project, register your fields, drop in one track() call. Your next deploy starts answering questions.

Start trackingRead the docs