# Tokenomix

> Tokenomix meters every LLM call your organization makes, charges it back to the team that made it, and screens prompts before they leave your network.

Tokenomix is a product of Polk Automation LLC.

## What it is

An LLM gateway that runs on your own infrastructure. Applications point at it
instead of at the model provider, and every request passes through five stages:
identity, quotas, screening, policy, audit.

## Key facts

- Deployment is on-premises or in your own cloud account. There is no hosted
  tier, no vendor tenant and no phone-home. Air-gapped installs work.
- API is OpenAI-compatible, including streaming, so integration is normally a
  base-URL change.
- It fronts both halves of an estate: hosted vendor APIs (OpenAI, Anthropic,
  Azure OpenAI, Google, Cohere) and models the customer runs themselves (vLLM,
  Ollama, or anything speaking the OpenAI-compatible API, on-prem or in their
  own VPC). The same policies, quotas and chargeback apply to both.
- Screening runs seven pattern detectors (jailbreak, PII, secrets, prompt
  injection, toxicity, topic restriction, custom regex) plus a fine-tuned
  classifier covering those categories and code injection, data exfiltration
  and encoding attacks.
- The classifier adds about 140 ms per request, a ceiling of roughly 8 requests
  per second on one GPU. Without a GPU, screening falls back to patterns only.
- Policy actions are block, warn, redact, log and allow. When several policies
  match, the strictest action wins, so a department can tighten a central rule
  but cannot weaken one.
- Quotas are enforced, not showback: tokens in, tokens out, total tokens,
  request count and cost in USD, over minute, hour, day or calendar month,
  scoped to a department, user, API key or provider. Each quota warns or blocks.
- Quotas are evaluated before screening, so an over-budget request costs no GPU
  time.
- Known limitation: token and cost totals are only knowable after the provider
  responds, so the request that crosses a cost limit completes and the next one
  is refused. Request counts are exact.
- Chargeback groups spend by user, model, provider, department or application,
  prices input and output tokens separately, and exports as CSV.
- Routing can pick the provider and model after screening has run, so a prompt
  containing PII or secrets can be pinned to a self-hosted model. It ships in
  shadow mode and is enforced only when switched on.
- Every request, verdict and cost figure is written to a database the customer
  runs, inside their own network.

## Pages

- [Home](https://tokenomix.ai/)
- [Platform overview](https://tokenomix.ai/platform)
- [Screening and policy](https://tokenomix.ai/platform/security)
- [Chargeback and quotas](https://tokenomix.ai/platform/cost)
- [Routing](https://tokenomix.ai/platform/routing)
- [Deployment](https://tokenomix.ai/deployment)
- [About](https://tokenomix.ai/about)
- [Request a demo](https://tokenomix.ai/demo)
- [Privacy](https://tokenomix.ai/privacy)
- [Terms](https://tokenomix.ai/terms)

## Contact

hello@tokenomix.ai