Zero-config cost & quality monitor for indie AI builders. One-line SDK, local dashboard, per-feature and per-customer attribution — and a budget kill-switch so an agent loop can never surprise you with the bill.
“We budgeted $4k/month for our agent. Three weeks later the bill was $11.2k — an uncapped recursive loop. The provider dashboard told us what we spent, not where.”
npx tokenwatch-sdk serve — one process, SQLite, no Docker, no Postgres, no ClickHouse. Your data stays on your machine.
One line for OpenAI or Anthropic, TypeScript or Python. No proxy in your request path — telemetry ships async on the side.
Cost by model, feature, and customer. Latency, error rates, daily trends. Set a budget — get webhook alerts at 80% and a hard stop at 100%.
import { wrapAnthropic, init } from 'tokenwatch-sdk';
const claude = wrapAnthropic(new Anthropic(), { feature: 'summarize', customerId: 'acme' });
init({ enforceBudget: true }); // over budget → BudgetExceededError, not a bigger bill
// That's it. Every call (streaming included) is now tracked:
// model, tokens, cost, latency, errors — attributed to features and customers.Your requests go straight to the provider. A monitoring tool should never be a single point of failure in your request path.
Cost per feature and per customer — know which features lose money and which customers are profitable, not just the total.
Set a monthly budget. At 80% your webhook fires; at 100% wrapped calls throw instead of spending. Agents can't out-loop you.
One process and a SQLite file. No ClickHouse + Postgres + Redis + S3 stack just to see your own spend.
Both SDKs included. The Python SDK has literally zero dependencies — standard library only.
Open source. Flat simple pricing for the upcoming cloud version — never per-seat enterprise games.
| TokenWatch | Helicone | Langfuse | Enterprise APMs | |
|---|---|---|---|---|
| Actively developed | ✓ | maintenance mode | ✓ (ClickHouse-owned) | ✓ |
| Request path | no proxy | proxy or async | async | async |
| Self-host stack | 1 process + SQLite | heavy | ClickHouse+PG+Redis+S3 | n/a |
| Budget kill-switch | ✓ built-in | limits | — | alerts only |
| Built for | solo devs / small teams | teams | teams | enterprises |
Helicone entered maintenance mode after its acquisition (March 2026). Langfuse is excellent for enterprise tracing — TokenWatch is for the rest of us.