margined 0.2.0

LLM unit economics: cost, margin, and profitability per customer and feature.

Margined Python SDK

LLM unit economics: cost, margin, and profitability per customer and feature.

Add one argument to your existing LLM calls. Get gross margin per customer, cost per feature, and the price you need to charge — without changing how your app works.

pip install margined

Quickstart

import margined
margined.init()  # reads MARGINED_API_KEY

response = margined.track(
    client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=[{"role": "user", "content": prompt}],
    ),
    user_id=current_user.id,
    feature="summarize_document",
)

track() returns the exact same response object, unchanged. Zero risk to existing code.

Auto-patch (zero call-site changes)

import margined
margined.init()
margined.patch_anthropic()   # sync + async clients
margined.patch_openai()      # sync + async clients

# Tell Margined who the current user is (e.g. Flask):
margined.set_context(
    user_id=lambda: g.current_user.id,
    feature=lambda: request.endpoint,
)

Every LLM call in your app — including calls made by third-party libraries — is now tracked.

Decorator + scoped identity

@margined.feature("research_agent")
async def run_research(user, query):
    async with margined.identify(user.id):
        return await client.messages.create(...)

Streaming

Streams report usage only at the end; wrap them and Margined records one event when the stream completes:

stream = margined.track_stream(
    client.messages.create(..., stream=True),
    user_id=user.id, feature="chat",
)
for event in stream:
    ...

For OpenAI, pass stream_options={"include_usage": True} so the final chunk carries usage. Auto-patched clients wrap streams automatically.

Agent runs

Group multi-step agent workflows under one run — the dashboard shows cost per run (p50/p90/p99), not just cost per call:

with margined.run(user_id=user.id, feature="research_agent") as r:
    step1 = client.messages.create(...)   # auto-tracked
    step2 = client.messages.create(...)
    r.tag({"steps": 2})

Guarantees

The SDK is built to the same standard as Sentry/PostHog-class telemetry SDKs:

• Fail-open. Tracking never raises into your application and never blocks a request thread. If Margined is down, your app is unaffected.
• Background delivery. Events batch in memory and flush every 5 seconds (or at 100 events) on a daemon thread, with retry + exponential backoff on transient failures.
• Bounded memory. The queue caps at 10,000 events and drops the oldest under sustained backpressure (drops are counted and reported).
• No double counting. Every event carries an idempotency key; a retried flush can never inflate your costs.
• Cache-aware pricing. Prompt-cache reads and writes are priced at each provider's actual discount rates, and OpenAI cached tokens are separated from the uncached input count.
• Local cost computation. Cost is computed from a bundled, versioned price table (margined.PRICES_VERSION) — no network round-trip on the hot path. The ingest API independently recomputes cost server-side, so a stale client table can't skew your dashboard.

Providers

Anthropic, OpenAI, Google Gemini, Groq, Mistral, DeepSeek, Together AI — with prefix matching for dated model snapshots and provider-prefixed IDs (anthropic.claude-…, openai/gpt-…).

Configuration

margined.init(
    api_key="mgd_...",       # or MARGINED_API_KEY
    endpoint="https://...",  # or MARGINED_ENDPOINT (self-hosting)
    sample_rate=1.0,         # record a fraction of events
    disabled=False,          # or MARGINED_DISABLED=1 (tests/CI)
    debug=False,             # or MARGINED_DEBUG=1
)

Serverless / short-lived processes

Call margined.flush() before the function exits (or margined.shutdown() to also stop the worker). The SDK also registers an atexit hook that drains the queue with a single best-effort attempt.

Dashboard

Events appear in your Margined dashboard in real time:

• Gross margin per customer — connect Stripe to see revenue vs. cost
• Cost per feature — which product surface burns your bill
• Pricing calculator — break-even and recommended prices from your real p50/p90/p99 usage
• Margin-at-risk alerts — email/Slack before a customer goes underwater