MODELS
Every model, one workspace.
You switch models without switching tools, permissions or memory.
A catalogue, not a vendor
Four providers are wired into the engine: Anthropic, OpenAI, Google and Mistral. Your workspace declares the models it allows, and names one default.
Two other roles are named along the way: the one that summarises long conversations, and the one that handles utility work. Summarising a thread does not need the most expensive model. An agent keeps the workspace default, or points at another model on the list.
Capabilities are not retyped
Context window, output ceiling and pricing come from a snapshot of the public catalogue shipped with the engine. Nobody retypes “200,000” into a config file, and two applications cannot drift on the same value.
A model whose capabilities cannot be resolved is refused at boot. Not in production, in front of a user.
A retired id breaks nothing
Models get renamed and retired several times a year. When the id stored on an agent is gone from the catalogue, the engine substitutes the closest live one: same provider first, then closest by price, then context window.
The pick is deterministic, so two identical workspaces land on the same model, and the substitution goes to the logs. On a transient stream error it is the call that is retried, with a growing backoff.
Usage is measured, so is the cache
Every call records its input and output tokens, its cache tokens, its tool calls and its duration. Cost is estimated from the model's own pricing.
When a model is not priced yet, the cost stays empty rather than wrong. An approximate invoice is worse than a blank field.
How it works
You decide to run your drafting agents on a cheaper model, and keep the strongest one for case analysis.
Your workspace picks its models from the catalogue and names one default.
An agent keeps that default, or points at another model on the list.
On every call the registry resolves the real id, the context window and the output ceiling.
If the id is gone, the closest model takes over and a warning goes to the logs.
Tokens, cache and estimated cost are recorded for the call.