Skip to main content
A model gateway standardizes access to AI models. An agent capability layer includes model access but also exposes data, tools, reusable workflows, discovery metadata, and controlled action or payment surfaces. The distinction matters because many agents need current evidence and external operations in addition to language generation.

​
Side-by-side comparison

​
Choose a model gateway when

A model gateway is usually sufficient when:
  • The application primarily generates or analyzes content.
  • External facts are already supplied by the application.
  • The team has separate tool and data integrations.
  • The workflow does not need reusable cross-tool Skills.
Start with AIsa Models if this describes the requirement. You can use the model gateway without adopting every other AIsa surface.

​
Choose a broader capability layer when

A broader layer is useful when:
  • The agent must retrieve current information before answering.
  • Several data providers would otherwise require separate integrations.
  • Multiple agent clients should share the same task instructions.
  • The agent needs machine-readable discovery of available capabilities.
  • The workflow may include controlled writes or billable runtime purchases.
Start with Capabilities by Goal or Capabilities by Interface.

​
They are not mutually exclusive

The model gateway is a capability inside the broader layer. An application can begin with model inference and add APIs or Skills only when a real workflow needs them. This incremental approach avoids unnecessary orchestration:
  1. Use a model for reasoning or generation.
  2. Add a direct API when current external data is required.
  3. Package repeated orchestration as a Skill.
  4. Add authenticated actions only when side effects are necessary.
  5. Add programmatic payment controls only for runtime purchasing requirements.

​
Evaluation questions

Before choosing an architecture, ask:
  • Does the application need one model provider, several models, or models plus external data?
  • Who should own endpoint sequencing: the application or a reusable Skill?
  • Are any operations writes rather than reads?
  • Does the agent need to discover capabilities dynamically?
  • What budget, retry, and audit rules apply?
For a full product-fit checklist, see When to use AIsa.