The LLM gateway is one API endpoint for all of your AI models. It runs on the Anchor Node. It holds every provider key on the server. It gives each person a different key. The software is LiteLLM.

​
Why it exists

AI models come from many providers. Each provider has its own API, its own key, and its own pricing. That causes three problems:
A model key belongs to one provider account. When people share one key, everyone shares one identity and one bill. You cannot tell who used what.
Point a tool at DeepSeek, MiniMax, OpenRouter and the rest directly and each needs its own endpoint and key. Add a provider and every tool must change.
Spend is scattered across provider consoles. There is no single view of who spent what on which model.
The gateway fixes all three: one endpoint for every provider, one key per person with its own budget, and every request logged against the key that made it.

​
What you get

One entry, all models

Point your tool at the gateway once. Reach every model behind it.

Your own key

Cancel one person’s key. The other keys continue to work.

Real keys stay on the server

A colleague never sees your DeepSeek key or your MiniMax key.

A cost limit for each person

The gateway stops a key when the person reaches the limit.

​
What it can do

The gateway is not a dumb pipe. It adds a layer of intelligence between your tool and the model providers. The pages in this section are split into two groups. For users — how to connect a tool, pick a model, and use your key. Start with Configure your agent. For the admin — how to add models, build pools, set fallbacks, watch usage, and manage access. Start with Models and pools.

​
The two addresses

The gateway has two addresses. They are not the same. The api address refuses everything except model traffic. These requests show the split: Ready to connect a tool? See Configure your agent for copy-paste settings.

​
Speed

The gateway adds a small delay: each request makes one extra hop through the Anchor Node.
LiteLLM itself adds almost nothing. The delay you see is distance — the second network hop — not software.