Trust model
Handlers run with the same trust as LLM-generated code (both are inside the sandbox). The handler’s value over a raw OpenAPI call is in what’s exposed, not where it runs:
- If a realm hides a method from its
apis.ymlallowlist but uses it inside a handler, the LLM cannot call the raw method directly. - If a realm exposes both the raw method and a wrapper, the LLM can call either — the wrapper is a recommendation, not a barrier. Skills (
SKILL.mdfiles) are the right way to make sure the LLM picks the wrapper.

