> ## Documentation Index
> Fetch the complete documentation index at: https://moxn.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Models, Reasoning Effort, and Cost

> Choose a model connection, primary model, and reasoning effort, and understand what drives usage

An agent runs on your own model provider account. You pick a model tier and a reasoning effort per agent, and those choices, together with how long conversations run, shape both quality and cost.

## Model connections

A model connection is an API key for your provider account. Add one under **Settings > Connections > Models**.

| | |
| - | - |
| **Providers** | OpenRouter and OpenAI |
| **Billing** | Your provider account is billed for all agent usage |
| **Key handling** | Encrypted and used only on the server. Agents never see the key. |
| **Multiple accounts** | Add several connections per provider, for example one per team or budget |
| **Rotation** | Replace a key to rotate it, or revoke it to cut off access |

Each agent uses one model connection. Subagents share their parent's connection.

## Primary model

| Model | Best for |
| - | - |
| **Astra** | Complex reasoning and coordination |
| **Terra** | Balanced execution |
| **Luna** | Lightweight, focused tasks |

Start with Terra for most channels. Choose Astra for channels that handle open-ended investigation or coordinate subagents. Choose Luna for narrow, high-volume work such as answering routine questions from the knowledge base.

## Reasoning effort

Reasoning effort sets how much the model thinks before it answers: **Low**, **Medium**, **High**, or **Extra high**. Higher effort generally gives better answers on hard problems and uses more tokens and time per reply. Lower effort keeps simple exchanges fast and inexpensive.

## Compaction threshold

The compaction threshold is the conversation size at which the agent summarizes older messages to make room. It applies to each session separately.

| | |
| - | - |
| **Default** | 64,000 tokens |
| **Range** | 8,192 to 200,000 tokens |

A lower threshold keeps each request smaller and compacts sooner. A higher threshold keeps more of the conversation verbatim before summarizing. The threshold is about conversation size; it is separate from the model's maximum context window and from spending limits.

## Subagents

With **Allow subagents** on, the agent can delegate parts of a request to helper agents. A typical setup is Astra coordinating several Terra helpers.

| Setting | What it controls |
| - | - |
| **Helper models** | Which models helpers may use |
| **Default helper model and effort** | What a helper uses unless the agent picks otherwise |
| **Max concurrent helpers** | 1 to 3 helpers running at once |
| **Model-request budget** | 4 to 128 model requests per turn |

Helpers share the parent's model connection, Moxn identity, connector grants, and files. One request budget and one four-minute execution deadline cover the parent and all of its helpers together.

## Keeping usage predictable

* **Match the model to the channel.** A Luna or Terra agent at Low or Medium effort is enough for most question-answering channels.
* **Use `task` for independent work.** A fresh session starts small instead of carrying the channel's full history. See [Working with the agent in Slack](/guides/agents/using-in-slack).
* **Tune the compaction threshold.** Lower it for busy channels where old conversation matters less.
* **Bound subagents.** Keep concurrency and the request budget low until you see the agent needs more.
* **Use a dedicated connection.** A separate key per agent or team makes usage easy to attribute in your provider's dashboard.

<Note>
  Spending limits live in your provider account. The compaction threshold and the subagent limits shape how much work a turn does, but they are safety limits, not a dollar cap.
</Note>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.