cognocient

A drop-in wrapper around the OpenAI and Anthropic Python SDKs that reports usage to Cognocient asynchronously, so you get live cost attribution without changing your base_url or routing traffic through a proxy.

Supported providers: OpenAI and Anthropic only (CognocientOpenAI and CognocientAnthropic). Have a Gemini, Mistral, Groq, Together AI, or Azure OpenAI key instead? See Which providers does this cover? below before you install — this package doesn't have a wrapper class for you yet.

pip install cognocient[openai]      # or cognocient[anthropic], or both

# Before
from openai import OpenAI
client = OpenAI(api_key="sk-...")

# After
from cognocient import CognocientOpenAI as OpenAI
client = OpenAI(
    api_key="sk-...",              # your own real OpenAI key, used exactly as before
    cognocient_key="sk-cog-...",   # auths the async usage report only — no traffic routes through Cognocient
)

client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "hello"}],
    cognocient_feature="support-bot",   # optional attribution — same field names the proxy accepts as X-Cost-* headers
)

Every method the real SDK exposes still works unchanged. This wrapper only intercepts chat.completions.create() (messages.create() for Anthropic) to time the call and report its usage after the fact; everything else is forwarded to the real client untouched.

What is cognocient_key, exactly?

It's just an authentication credential — the same kind of API key you'd generate for the proxy — but here it does one thing only: it lets the wrapper's background reporting call prove to Cognocient's ingestion API which account a usage report belongs to. It has nothing to do with your actual OpenAI/Anthropic request. That request still goes straight to api.openai.com / api.anthropic.com with your own real provider key, exactly like it would with no Cognocient wrapper installed at all — no base_url change, no network hop, no dependency on Cognocient being up. The wrapper just fires a second, separate, best-effort call afterward to log what happened, off the critical path, on a background thread. Losing that reporting call (Cognocient down, key wrong, network blip) never affects your real API call — see "Reliability" below.

Which providers does this cover?

Only OpenAI and Anthropic. There is no CognocientGemini, CognocientMistral, CognocientGroq, CognocientTogether, or CognocientAzureOpenAI class in this package today — the Cognocient proxy covers all seven (OpenAI, Anthropic, Google Gemini, Mistral, Groq, Together AI, Azure OpenAI) because it works at the HTTP layer, not by wrapping each provider's SDK object one at a time.

If your key is for one of the five providers this wrapper doesn't cover:

1. Use the proxy instead — it's provider-agnostic, covers all seven providers, and adds pre-call budget enforcement this wrapper never provides for any provider.
2. Use the CSV/OTel importer if you want zero live network calls to Cognocient from your runtime — works for any provider, historical-only.

Don't try passing Groq's or Together AI's OpenAI-compatible base_url into CognocientOpenAI — it will run without erroring, but the reported provider field is hardcoded to "openai", so your dashboard will mislabel that spend. Not a supported path.

This is one of three ways to see your Cognocient dashboard

Security — read this before you decide

This wrapper is not more secure than the proxy. It is a different tradeoff, not a strictly better one.

With the proxy, your real provider API key lives server-side, under Cognocient's control, in one place. With this wrapper, your real provider key stays in your own application process, exactly as it does today without Cognocient at all — the wrapper calls the provider directly, using your key, inside your runtime. Some security teams prefer that (no third-party network hop in the request path); others are less comfortable with third-party code executing inside their process with key access. Both are reasonable positions. We're not going to tell you this "removes a security roadblock" — it trades one shape of exposure for a different one.

What this wrapper honestly gives you over the proxy:

• Zero added request latency. Reporting happens after your real call already returned, on a background thread, off the critical path.
• Zero risk of a Cognocient outage affecting your production call. If Cognocient's ingestion API is down or unreachable, your call to OpenAI/Anthropic still completes normally — see "Reliability" below.

What you give up versus the proxy: pre-call enforcement. Because Cognocient only hears about a call after it already happened, budgets configured in Cognocient cannot block or degrade a call made through this wrapper before it fires. The dashboard will say so explicitly for any account using this path.

Reliability

Reporting is fire-and-forget on a background thread with a bounded local queue, flushed every few seconds or every 50 calls, whichever comes first. If the ingestion API is slow, down, or unreachable:

• Your real provider call is completely unaffected — it already happened before reporting was attempted.
• No exception is ever raised into your code from a reporting failure.
• No retry loop that could pile up work in your process — a failed batch is dropped and logged locally at DEBUG level via the cognocient logger, not retried.

See tests/test_reporter_failure_isolation.py for a test that simulates an unreachable ingestion endpoint and asserts the real call still completes normally.

Known limitation: streaming isn't reported yet

stream=True calls are passed through to the real SDK completely unmodified — your application behaves identically — but are not currently reported to Cognocient. Usage totals aren't available until a stream completes, and reliably capturing them requires wrapping the stream iterator itself, which this version doesn't do. If most of your traffic streams, this wrapper will under-report your usage today. Use the proxy or the CSV/OTel importer if that matters for your evaluation.

Attribution fields

Same field names the proxy accepts as X-Cost-* headers, passed as keyword arguments instead:

Development

pip install -e ".[dev]"
pytest
python benchmark/benchmark_wrapper_overhead.py