valuemaxx 0.1.3

AI Margin Intelligence — Python SDK + optional CLI/backend. One-line init() that fails open.

valuemaxx (Python SDK)

Know what each AI agent actually costs you — correctly — and what it earned, per outcome, with confidence.

The Python SDK for valuemaxx — AI margin intelligence for teams that build AI agents. One line captures the correct cost of every LLM call (off the hot path, never throwing into your app), so valuemaxx can bind it to the business outcome each agent run produced.

Install

pip install valuemaxx

Use

import valuemaxx

valuemaxx.init()   # zero-code cost capture starts here

That's it. init() instruments the openai / anthropic clients (and frameworks that use them — LangChain, LlamaIndex, CrewAI) at the transport layer, capturing correct token/cost per call: streaming usage taken from terminal values (never delta-summed), cache token classes priced separately, retries counted. It fails open — internal errors are caught and logged, never propagated into your call path, and it adds <1ms.

Attribute a run to an outcome

Wrap a unit of work so its cost can later be tied to a business outcome:

with valuemaxx.run(customer_id=user.id):
    response = client.messages.create(...)   # cost captured AND tagged to this run

The entity IDs you pass are the natural keys valuemaxx uses to bind delayed/out-of-process outcomes (a deal closing days later, a ticket resolving) back to the run that caused them — with an honest confidence label on every link.

What's captured (and what isn't)

• Captured: model, token counts (input / cache-read / cache-write / output / reasoning), computed cost with provenance (measured → reconciled to the provider invoice later), latency, the run context.
• Off by default: prompt/response content (PII-sensitive). Enable it only for the eval/replay features, on self-hosted deployments.

The honesty model

Every number valuemaxx produces carries its trustworthiness — cost provenance (measured/estimated/allocated/provider_reconciled), and once joined to outcomes, a binding tier and signal class. A rollup always shows the least-trusted of its parts; estimates never render as billed.

Self-hosting

The SDK ships telemetry over OTLP/HTTP to a valuemaxx backend you run (a container + Postgres; SQLite for local dev). The SDK itself holds nothing and needs no database. See the main repo for the backend, the TypeScript SDK (npm i valuemaxx), and the onboarding flow that wires outcomes for you.

License

Apache-2.0