A Squad reaches its verdict by calling a language model many times. The good news is that one is included — the platform provides a default model, so a Squad runs with nothing connected. You only come to this page when you want to use a better or specific model, route different jobs to different models, or cap what you spend.
You manage all of it on the Agent Models page, which has two halves: Connections (which models you can use) and Routes (which model does which job).
Included, or bring your own#
The distinction that matters most is free-and-included versus better-and-yours.
Three ways to connect a model#
Connections come in three kinds, mirroring how you connect an exchange.
| Kind | What it is | Cost | When to use it |
|---|---|---|---|
| Platform | The models the platform provides, including the free default. | Included | The simplest start — nothing to set up. |
| Direct | Your own provider API key, encrypted at rest. | You pay the provider | A stronger model, or one not in the platform pool. |
| Remote | A model you host yourself, reached through a gateway. | Your hardware | Full control and privacy — keys and prompts never leave your machine. |
Supported providers include Anthropic (Claude), OpenAI (GPT), Voyage (for the embeddings behind semantic search), Ollama for local models, and any OpenAI-compatible endpoint (vLLM, LM-Studio, llama.cpp).
A Direct connection takes your own API key and encrypts it at rest. A Remote connection routes through the self-hosted gateway, so your key and the Squad's prompts stay on hardware you control — the same self-custody idea as hosting your own exchange keys.
Running Squads on a local model through Ollama or a remote gateway turns the per-call cost to zero. It is the natural setup for a Squad that deliberates around the clock on a cadence — the included model gets you started, a self-hosted one keeps a round-the-clock Squad cheap.
Routing jobs to models#
Not every step of a deliberation needs the same muscle. Writing a specialist's long-form analysis is a different job from a one-line risk verdict, and you can send each to a different model. The platform groups the work into task classes, and a route points each class at a provider and model.
| Task class | Used for |
|---|---|
| Deep reasoning | Long-form analysis — the specialists and managers |
| Quick judgment | Short calls — risk verdicts, classifications |
| Data extraction | Pulling numbers, indicators, sentiment |
| Summarization | Compressing prior turns into briefs |
| Tool calling | Invoking tools and external APIs |
| Embedding | Vector embeddings for semantic search |
This lets you put a strong, expensive model on the deep reasoning and a cheap, fast one on the quick judgments — most of a Squad's calls are not the hard ones, so the savings are real. Each route has a Test button that fires a one-token call to confirm it is wired up.
The fallback chain#
You do not have to set a route for everything. A Squad resolves each task class through a chain, taking the first one that is set.
When you create a Squad, the Model routes per task class (advanced) section lets you give that Squad its own routes. Leave it alone and the Squad inherits your defaults, which inherit the platform default.
Capping what you spend#
Because a Squad on a cadence deliberates on its own schedule, it can also spend on its own schedule. A daily spend cap bounds the total tokens your Squads may use in a day; once the cap is hit, deliberations stop until the next day rather than running up an open-ended bill. Raise or lower it as your usage and cadence change.
A Squad on a fast decision interval calling a top-tier model on every task class can spend quickly. Set the cap before you put a Squad on a cadence, not after — and consider a local or remote model for anything that runs continuously.