A fixed cost per unit. Your stack stays yours.

Nothing in your request path until you sign. Then a small change in each SDK, and your own keys stay as the fallback.

Before you sign

We read metadata. Never prompts.

The X-ray
Usage exports
The shadow period
Call metadata

Tokens, model, latency, retries, and the unit. Our exporter is open source, and you scope its key.

When the guarantee starts

What changes in your code

Base URLthe provider's→our endpoint
Keyyours→ours
Each call+a unit ID header
Each SDK+a few lines of fallback

Your prompts stay exactly as they are.

What stays yours

Your keys, your prompts, your models

Your keys

Stay in your client as the fallback. If we're down, calls go straight to the provider.

Your prompts

Passed through unchanged. Never stored.

Your models

Nothing changes without your approval, tested on your own test set.

Providers
  • OpenAI
  • Anthropic
  • Google Gemini
Features
  • Streaming
  • Tool calls
  • Structured outputs
  • Embeddings
  • Prompt caching

Book 15 minutes