useleak 0.1.0

Drop-in cost tracking for the Anthropic Python SDK — see exactly where your LLM spend goes.

useleak (python)

Drop-in cost tracking for the Anthropic Python SDK. See exactly where your LLM spend goes — broken down by feature, user, and model.

Install

pip install useleak

Usage

Get a workspace SDK key from Settings → Workspace SDK key in your Leak dashboard. Set it as LEAK_KEY in your env (alongside ANTHROPIC_API_KEY).

import os
from useleak import Anthropic

client = Anthropic()  # picks up ANTHROPIC_API_KEY + LEAK_KEY from env

response = client.messages.create(
    model="claude-opus-4-7",
    feature="icp-scoring",     # ← the one extra param Leak needs
    user="user_123",            # optional but lets you group by customer
    max_tokens=1024,
    messages=[{"role": "user", "content": "..."}],
)

print(response.content[0].text)

That's it. Every call gets tracked to your Leak dashboard within a few seconds: model, tokens in/out, cost, feature, user, request ID.

API

useleak.Anthropic accepts every kwarg the official anthropic.Anthropic accepts, plus two of its own:

Parameter	Default	Description
leak_key	LEAK_KEY env	Your workspace SDK key (lk_live_...)
leak_url	https://useleak.cloud/api/ingest	Ingest endpoint, override only if self-hosting

client.messages.create(...) accepts every kwarg the Anthropic SDK accepts, plus three of its own:

Parameter	Required	Description
feature	recommended	The feature in your app that triggered this call ("icp-scoring", "playground", etc.)
user	optional	An end-user identifier — lets you see who is burning tokens
api_key_label	optional	Free-form label like "Production" / "Staging"

Tracking is async and non-blocking. If useleak.cloud is down, your Anthropic call still succeeds — Leak failures never propagate.

What it tracks

• model
• input tokens / output tokens
• USD cost (computed locally from the Anthropic price table)
• feature, user, api_key_label (your tags)
• Anthropic request ID

It does not send prompts, responses, or any message content.

License

MIT