Skip to main content
Every turn runs on one model, from one credential. Both are named by the agent that is running:
model is required. Alongside it an agent names exactly one of two things, and neither is optional: Naming both is refused rather than resolved: an agent that names a provider while running on somebody else’s credential lies to whoever reads it, and there is no reading of it that bills the right account. Naming neither is refused too, in a sentence saying which of the two to write. Both are checked when the run starts, not at the moment agent() runs.

​
”Harness decides” in the composer

The chip at the bottom right of the composer is the model for the messages you send in this chat.
The Splox composer menu with Model set to Harness decides

The composer menu: Model, Effort, and Advanced

Harness decides is the default, and it means exactly what it says: no selection is stored for this chat, so the model is the one the running agent declared. Under the hood the chat’s composer state carries "selection": null. Open Model and you get the catalogue this account may use, with Harness decides checked at the top:
The Splox model picker

The model list. Picking one overrides the harness for this chat.

Picking one writes a selection against the chat — the endpoint and the model — and picking Harness decides again clears it. Nothing about the harness changes either way; the selection lives with the conversation.

​
Effort

Effort is the reasoning ladder, and it only has rungs once a model is chosen: with Harness decides the menu offers Default alone, because the ladder is a property of the model and no model has been named yet.
The Splox effort menu showing Default, Off, Low, Medium, High, Max

With a model chosen, the effort ladder is that model's own

The rungs are off, low, medium, high, max, and each model advertises which of them it has and which is its default. A model whose catalogue row has no such rung refuses the turn rather than quietly running at another one. In a harness the same idea is the agent’s reasoning= field; both starter agents write reasoning="max".

​
Choosing a model per turn

The declaration is the answer to “which model”, asked once when the run starts. hooks/model.py is the same question asked again on every turn:
programs/splox/hooks/model.py
Answer a bare string to change the model alone, or an object to move the turn to another provider — which is the only way a model name from a different vendor means anything:
The opening of a conversation is where the plan is made and where a wrong plan costs the most, so it is the turn worth spending a slower model on. model.choose runs after context.build, so it overrides a model a context named, and it is only consulted when the hook actually returns a name. See Hooks.

​
Endpoints, and your own keys

An endpoint is a credential plus the base URL it talks to. Every account sees the platform’s own four: Writing provider="anthropic" on an agent is how you ask for the platform’s Anthropic credential; you never handle its key. To run on your own, add it under Connections in the sidebar — “Connect a provider API key, ChatGPT account, or custom OpenAI-compatible endpoint” — and then name the endpoint on the agent by its id:
The OpenAI adapter takes any OpenAI-compatible base URL, which is what makes a local model or a third-party gateway reachable the same way. List the endpoints an account may use, and the models behind one, with /v2/llm-endpoints.
The voice leg has its own two fields — voice_llm_model and voice_llm_endpoint_id — used when an agent is invoked with a destination number. Most authors leave them out and reuse one endpoint for both legs.

Agents

Every generation field: reasoning, output cap, temperature, caching, modalities.

LLM endpoints API

Discovering the endpoints and models available to a key.