Before relying on a stop control, trace the model calls it can actually stop. A gateway can govern traffic routed through it. A direct call to a model provider is outside that enforcement path.
That distinction matters when you move an agent from a local experiment into a product people use.
Start With The Agent, Not Another Execution Loop
If you’re building a product that calls models, rebuilding the same agent execution loop can burn an afternoon before you reach the product-specific work.
Kimss Forge is Kimss AI’s MIT open-source agent harness: local, free, and usable without a Kimss account. Start with a few lines and your model endpoint rather than building the harness yourself.
pip install kimss-forge
The project’s BENCHMARK.md is the reference for the “about 80% less app code than the hard way” comparison—not a guarantee for every application.
A Warning Is Not An Enforcement Boundary
A signal tells you something happened. A policy defines what should be allowed. Enforcement needs a point where the action can actually be stopped.
Forge keeps those layers distinct:
- Local execution: build and test without a Kimss account.
- Soft Authority Boundary warning: a sensitive tool firing without a production gateway produces a terminal warning. It never blocks local execution.
- Gateway enforcement: controls apply when traffic passes through the Kimss gateway.
The warning is useful feedback. It is not proof that the tool action was blocked or that local execution has production protections.
Attach The Same Agent When You Need A Control Point
Forge’s production path is to add gateway="kimss" to the same agent.
On the free Developer plan, that gateway path adds identity on model calls and the Kill Switch. The Kill Switch severs access at the Kimss gateway for routed traffic; it does not terminate customer processes that never call the gateway.
For an existing OpenAI-compatible client, the routing change is its base_url setting:
base_url="https://api.kimss.ai"
That line shows the endpoint swap, not the complete authentication setup. You bring your own model infrastructure; Kimss is a model-agnostic API gateway, not a model host or compute reseller.
Usage is metered in governed requests, not model compute sold by Kimss.
Keep The Capability Layers Separate
You do not need to adopt every production capability to start shipping.
- Forge: the local, MIT-licensed starting point.
- Free gateway path: identity and Kill Switch for routed model traffic.
- Paid capabilities: Authority Boundary, SSO, and SCIM, available later on the same path without rewriting the agent.
Do not confuse the local Authority Boundary warning with paid Authority Boundary enforcement. They serve different purposes.
Check The Route Before Trusting The Stop Button
Before you rely on gateway controls:
- Trace each model call’s destination.
- Identify any calls that go directly to a provider.
- Verify which requests actually pass through the gateway.
- Keep bypassing calls outside your claims about gateway enforcement.
You can keep local development simple without pretending local execution is governed production traffic.
Next step: run pip install kimss-forge. Build locally first, then attach the gateway when you need a control point for production model calls.
Top comments (0)