If you were using Helicone to track your LLM costs, you've probably seen the news: after the Mintlify acquisition in March, it's officially in maintenance mode. Feature development has stopped. 16,000+ organizations are quietly looking around. Langfuse — the other indie-friendly option — was acquired by ClickHouse in January, and self-hosting it means running ClickHouse + Postgres + Redis + S3. To look at your own API bill.
Meanwhile the problem is getting worse, not better. We're all running agents now, and agents have a special talent: an uncapped recursive loop can turn a $4k/month budget into an $11.2k bill in three weeks (real story). The provider dashboards tell you what you spent. Not where, not which feature, not which customer.
So I built TokenWatch — the tool I wanted as a solo AI builder:
npx tokenwatch-sdk serve # dashboard on localhost:4318. That's the whole setup.
js
import { wrapAnthropic, init } from 'tokenwatch-sdk';
const claude = wrapAnthropic(new Anthropic(), { feature: 'summarize', customerId: 'acme' });
init({ enforceBudget: true });
Every call — streaming included — is now tracked: model, tokens, cost, latency, errors, attributed to features and customers.
Design decisions (a.k.a. my complaints about the incumbents)
No proxy in your request path. Your calls go straight to OpenAI/Anthropic; telemetry ships async on the side. A monitoring tool should never be the reason your product is down.
One process, SQLite, zero native deps. It uses Node's built-in node:sqlite. No Docker compose with four services. Your usage data stays on your machine.
A budget kill-switch, not just a budget chart. Set a monthly budget: at 80% your webhook fires, at 100% wrapped calls throw BudgetExceededError instead of spending more. Watching a dashboard doesn't stop an agent loop at 3am — an exception does.
Margin attribution, not traces. Tracing UIs are built for debugging. Most of the time I have a simpler question: which feature is losing money and which customer is profitable? Cost by feature and by customer is the default view, not a saved query.
Python SDK with literally zero dependencies. Standard library only. wrap_openai(client) and you're done.
What it's not
It's not a tracing platform, it's not an eval suite, and it won't replace Langfuse for a 50-person team that lives in traces. It's the 80% tool for the solo builder and small team: where is the money going, is quality degrading, and stop the bleeding automatically.
It's MIT-licensed and v0.1 — built in public, partly with AI agents (Claude Code wrote a lot of it, which felt appropriately recursive for a tool that monitors AI spend). Feedback, issues, and brutal honesty welcome.
GitHub: https://github.com/jkhusanovpn/tokenwatch
What's your current setup for tracking LLM costs — and has an agent ever surprised you with a bill?