FiscaLLM

See the AI cost of every single user. FiscaLLM adds up what each user costs you in AI and sets it next to what they pay you, so you can see who is profitable and who is losing you money.

Open source (MIT) and free. Use the hosted service at fiscallm.com (free under fair use, best effort), or run your own.

Live demo (no sign-up) · Docs · How it works · What it covers · Accuracy · Roadmap (board) · Changelog · Self-hosting · Contributing

Sample data from the demo, not a real customer.

Get started in three steps

1. Create an app at fiscallm.com (or your own copy) and copy the ingest key.

2. Add one line after each AI call (it reads only the model name and token counts, never prompts or answers):

   import { createMeter } from "fiscallm"; // npm install fiscallm   (Python: pip install fiscallm)
   const meter = createMeter({ apiKey: process.env.FISCALLM_KEY, waitUntil });
   
   const res = await openai.chat.completions.create({ model, messages });
   meter.trackOpenAI(res, { userId, feature: "chat" });
   

3. Connect revenue: press Connect RevenueCat (one click: adds the webhook and imports past subscribers), paste a Stripe webhook URL, or send payments yourself with the revenue API.

No AI key handy? FISCALLM_KEY=fsk_... node examples/send-sample-data.mjs fills a test app with sample users so you can see it working.

The problem

AI apps charge a flat price but pay a variable cost. One user costs $0.20 a month in AI, another costs $25. Your AI provider's bill shows tokens, your subscription tool shows payments, and neither shows the user.

How it works

import { createMeter } from "fiscallm";

const meter = createMeter({ apiKey: process.env.FISCALLM_KEY, waitUntil });

const res = await openai.chat.completions.create({ model, messages });
meter.trackOpenAI(res, { userId, feature: "chat" });

from fiscallm import Meter

meter = Meter(api_key=os.environ["FISCALLM_KEY"])
res = client.chat.completions.create(model=model, messages=messages)
meter.track_openai(res, user_id=user_id, feature="chat")

• The SDK reads only the model name and token counts from the response. Never prompts or answers.
• Events are batched and sent in the background. Tracking never throws and never blocks your response.
• Each event has a client-generated UUID, so retries cannot double count.
• Revenue comes from RevenueCat (one-click OAuth connect, or a webhook you add by hand), or from your own code with trackRevenue / track_revenue.
• Both are joined on the user ID. RevenueCat aliases are resolved, so anonymous and logged-in users merge.
• Budgets and alerts: email, Slack or signed webhooks when a user goes over budget, when a paying user loses money, or when one user burns a huge number of tokens in 5 minutes (a runaway loop).
• Jobs (pay-per-job apps, e.g. credits for a video render): tag AI calls with a jobId and report each job as pending, captured or refunded with meter.trackJob / track_job. The Jobs page shows the AI cost of refunded jobs and of jobs that never resolved, per user.
• Spending caps: mark a per-user budget as Block and call meter.check(userId) before an AI call (TypeScript, Python). It says no once the user reaches the amount, If FiscaLLM is slow or down, a user it already refused stays refused; anyone else is allowed by default (or refused, with checkOnError: "deny"). An optional local limit counts each user's calls and tokens in your process, so a runaway loop is stopped even with no network.

Cost is priced on arrival: exact costUsd if given, otherwise per-token (input, output, cache read and cache write priced separately) or per-unit (images, audio, video) from a price table. Unknown models are flagged as unpriced, never counted as zero.

Net revenue is price × (1 − tax − store commission).

Layout

Path	What it is
apps/web	Next.js site and dashboard
apps/api	Hono API: ingest, RevenueCat webhook and OAuth callback, revenue API, cron jobs
packages/shared	Drizzle schema, pricing, metrics, alerts
packages/sdk	TypeScript SDK (fiscallm on npm)
packages/sdk-python	Python SDK (fiscallm on PyPI)

Stack: TypeScript, Next.js, Hono, Postgres (Supabase) with Drizzle, hosted on Vercel.

Limits

• It needs a backend between your app and the AI provider. Apps that call a provider straight from the phone are not covered yet.
• Revenue comes from RevenueCat, Stripe (beta, USD only) or the revenue API. Direct App Store and Google Play connections are not built yet.

Try it in ten minutes

• No AI key needed: FISCALLM_KEY=fsk_... node examples/send-sample-data.mjs fills the dashboard with sample users (one heavy user who loses money).
• Copy-paste starting points for Next.js and FastAPI, see examples.
• Run your own copy: (create a Supabase project first, see SELF-HOSTING.md).

Development

pnpm install
pnpm dev:web   # http://localhost:3000
pnpm dev:api   # http://localhost:8787/api/health
pnpm test

Copy .env.example to apps/web/.env.local and packages/shared/.env and fill in the values.

pnpm db:generate
pnpm db:migrate

Contributing

Issues and pull requests are welcome. Read CONTRIBUTING.md first (it includes a short contributor agreement). The plan, including what we only build if there is demand, is in ROADMAP.md, with a public board. Vote with a 👍 on roadmap issues or share ideas in Discussions.

Where help is most wanted: prices for more models and providers (data only, no code knowledge needed), connectors for other billing systems (Adapty, Superwall), examples for AI frameworks, and a comparison of computed costs with a real provider invoice. See the good first issues.

Built by

Built and maintained by Mayur as an independent, open-source side project. Questions, bugs and ideas: open an issue or email [email protected].

Contributors

Thank you to everyone who helps. Contribute: a good place to start is an issue labelled good first issue.

License

MIT. See LICENSE.