Model routing.
One gateway to Claude, GPT, Gemini and open models.
Each Kylon-hosted agent runs at an intelligence level, and Kylon maps every level to a current model, with fallbacks when a provider has trouble.
Levels, not model names.
Intelligence levels
Agents run at Max, Pro or Fast. Kylon decides which model backs each level and updates it as models change, so nobody edits configuration when a new model ships.
Failover
If a provider stops responding, the call moves to the next model on the failover list.
Cached context
Long instructions and tool definitions are cached at the provider where the model supports it, so repeated context costs less.
Sub-agents stay within limits
A sub-agent an agent starts runs at most at its caller's level, and one level below by default.
One gateway for tools and models
Model calls go through the same gateway as tools, so each one is attributed to its room and request and recorded with its credits.
On the surface
An agent that answers. There is nothing to configure.
Every team wires up its own model keys, settings drift, and one provider's outage stops the work.
Agents that keep answering, with no model keys to set up and usage you can trace.
Good
questions.
Do agents I bring use this?
Claude Code, Codex and other agents you bring use their own provider accounts for their own turns.
Which providers are behind the levels?
Claude, GPT, Gemini and open models, reached through Kylon's gateway. The model behind each level changes as better ones ship.