myGate

myGate is a cost-control gate for nested AI agents: it denies known superheavy model launches before they start and notifies on runtime-only ones, preventing expensive fan-outs from happening silently. It ships as a Python CLI and PyPI package.

The problem

Nested agent frameworks — where one LLM spawns sub-agents that may themselves spawn further calls — can trigger model launches that each cost dollars per call. Without a gate, a misconfigured agent tree can rack up significant spend in minutes, and the operator only finds out from the billing dashboard after the fact.

The approach

myGate maintains a registry of known superheavy models and intercepts launch requests before they reach the provider. Known heavies are denied outright; models only discoverable at runtime (new releases, dynamic routing) trigger a notification instead of a silent pass-through. The 9-test suite and 8 documentation pages cover the deny and notify paths. Published on PyPI under MIT at version 0.4.0 with CI, pre-commit hooks, and typed source.

What it demonstrates

A stage-3 published gate with CI, pre-commit, and typed source shows cost-awareness turned into deployable infrastructure — the difference between "we should watch spending" and "a program enforces it before the request leaves the host." At this maturity the surface is small and the registry is the value; the engineering discipline is present but the tool is still early.