The Shoreheld gateway

A gateway in your own cloud account sits between your engineers' tools and the models they call.

What it checks

  • Each prompt is checked for secrets and customer data before it leaves your network.
  • Each response is checked on the way back, and risky commands are blocked.
  • Keys last one hour.
  • Every request is written to your own logs.

Where it runs

The gateway is open-source LiteLLM, deployed in your cloud account under your keys, with the configuration in your repository. Shoreheld has no standing access to it.

If Shoreheld is unavailable, the gateway sends work to the labs you already use, so engineers keep working.

What does not go through it

Features that run on a lab's own servers, such as cloud sessions and hosted connectors, cannot pass through a gateway in your network. You decide which of those to leave switched on.

The pilot

A pilot answers one question: how much of your work can an open model do, and at what cost?

  1. A task is a change your engineers have already merged, with its ticket and its tests.
  2. We reset the repository to just before the change and give the model the ticket.
  3. The task passes if the change's own tests pass. The model never sees those tests.
  4. Every task runs on both the open model and the model you use today.
  5. An engineer reviews a sample of passing changes without knowing which model wrote them.

You get the pass rate for each model and the cost per solved task.

Ask about a pilot

Tell us which tools your engineers use and roughly how much you spend on models each month.