apipurse-sdk 0.1.0

APIPurse real-time cost tracking SDK for Python — automatic OpenAI/Anthropic wrappers plus a manual track_completion() client. Not a network proxy: nothing routes through APIPurse's servers.

APIPurse SDK — Python

Python port of sdk/apipurse-sdk (the TypeScript reference implementation). Dependency-free (standard library only) client for /api/track/completion — the real-time tracking endpoint. Verified end-to-end in this repo's sandbox against fake OpenAI/Anthropic-shaped clients (streaming + non-streaming).

Two integration styles, both call the same endpoint under the hood:

• Automatic (wrap_openai() / wrap_anthropic()) — wrap your existing client once at setup; every subsequent .create() call is tracked with zero changes at each call site. Works with both sync and async clients (OpenAI/AsyncOpenAI, Anthropic/AsyncAnthropic) — detected automatically.
• Manual (APIPurse.track_completion()) — call it yourself right after a completion.

Neither is a network proxy: no request or response data is ever routed through APIPurse's servers. Only the model ID and token counts are sent.

Install

Not yet published to PyPI. Copy the apipurse/ folder into your project, or:

pip install -e /path/to/sdk-python

Usage

import os
from openai import OpenAI
from apipurse import APIPurse, wrap_openai

apipurse = APIPurse(
    ingest_token=os.environ["APIPURSE_INGEST_TOKEN"],
    provider_id=os.environ["APIPURSE_PROVIDER_ID"],
    base_url="https://your-apipurse-domain.com",
)

client = wrap_openai(OpenAI(), apipurse)

# Tracked automatically:
completion = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
)

For Anthropic:

from anthropic import Anthropic
from apipurse import wrap_anthropic

client = wrap_anthropic(Anthropic(), apipurse)
message = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)

Manual path (any provider, any call shape):

result = apipurse.track_completion(
    model="gpt-4o", input_tokens=120, output_tokens=48,
)

Publishing to PyPI

Ships as plain source. To publish: add a [project] version bump workflow, then python -m build && twine upload dist/*. Needs a PyPI account/org you control — not done here.