Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.opper.ai/llms.txt

Use this file to discover all available pages before exploring further.

Opper is one API in front of every major AI model, wrapped in a layer that governs and records every call passing through it. You write ordinary model calls. Opper routes each one to a model, applies your rules, and records what happened. Two ideas cover almost everything: the two planes, and how you organize work into projects.

​
Two planes

The AI Gateway is the request path. Your app calls it, it picks a model, runs the request, and returns the result. One API for Chat, JSON, and Realtime, in front of 300+ models. The Control Plane wraps around it. It’s where you decide which models can run, what content gets blocked, how responses are scored, and what you spend. The relationship is the whole point: you configure rules in the Control Plane, and the Gateway enforces them on every call, automatically. Your application code stays clean, and the policy lives outside it. Move a function to a cheaper model, block a category of content, or cap monthly spend by editing a rule, never by shipping new code.

​
Projects

Your account is an organization. Inside it, each app or environment is a project with its own API key and its own isolated data. Control Plane rules attach at one of these two levels. An org rule applies everywhere; a project rule narrows it. You can tighten at the project level but never loosen what the org allows, so a policy like “EU providers only”, set once at the org, holds across every project underneath it.

​
What happens on a call

Every call runs the same path, and you can replay all of it afterward:
  1. The Gateway identifies your project from the API key.
  2. Guard checks the input, Comply checks the model is allowed, and Route fills in a default model if you didn’t name one.
  3. The model runs.
  4. Guard checks the output, and Observe scores it.
  5. The call is recorded. By default that’s metadata only (model, cost, latency); with retention on, it’s the full trace.
A trace is the full tree behind one request: the model call, any tool calls, and every rule that fired. By default Opper keeps only metadata, so to store and open full traces (with inputs and outputs) you add a Comply retention rule. That’s also what lets Observe score a call, since it needs the content to judge.

​
Where to go next

AI Gateway

One API, 300+ models, routing, and EU enforcement. The request path in detail.

Control Plane

The five tools that govern, observe, and improve what’s running in production.