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

# Choose an agent

> Choose what a new chat runs on, select a model, and check how its usage is paid.

Use the agent picker in the composer to choose what a new chat runs on. Make your choice before sending its first message.

| Option | How to use it |
| - | - |
| **Cohesive** | Work with the built-in agent using platform credits and its reasoning controls. |
| **Claude** | Choose an available Claude model. Connect your [Claude subscription](/product/account/connect-claude) or an Anthropic API key to use your own account. |
| **ChatGPT** | Choose an available OpenAI model. Connect your [ChatGPT subscription](/product/account/connect-chatgpt) or an OpenAI API key to use your own account. |
| **OpenCode** | Choose an available OpenAI or Anthropic model using an eligible API key. Subscription credentials aren't used. |

The picker shows which options are available to you and explains unavailable choices. If agent choice isn't enabled for your organization, you continue with Cohesive.

## Start a conversation with your choice

<Steps>
  <Step title="Open a new chat or session">
    Start from Home, open Quick Chat, or create a new session on a Canvas.
  </Step>

  <Step title="Choose the agent and model">
    Open the agent selector in the composer. For options with a model menu, choose a model and check its funding label.
  </Step>

  <Step title="Set permissions and send">
    Choose a [permission mode](/product/agent/permissions), add your context, and send the first message.
  </Step>
</Steps>

## What can change during a chat

The agent choice is fixed after the first message. Start a new conversation to switch from Cohesive to Claude or ChatGPT, or between agent options.

The model can change for later messages within that agent's available choices. OpenCode conversations stay with the provider they started on; start a new chat to change providers.

Your [permission mode](/product/agent/permissions) can also change for the next message.

## Set thinking effort

Where the composer shows a thinking-effort control, choose the level that suits the task:

* **Low.** Fetch information, place content, or make a quick edit.
* **Standard.** Create diagrams, presentations, and other structured content for everyday work.
* **High.** Work through complex Canvas tasks or research that needs more depth.

## Check how the turn is paid

The model menu shows **your subscription**, **connected key**, or **platform credits**. A subscription connection applies only to compatible agent options. See [Model providers](/product/account/model-providers) for connection precedence and credit accounting.

## Keep the context with the work

A new conversation has its own history. On a Canvas, select the source cards and results you want the new agent to use. Ask for a specific output and check the result against those sources.

Different agent options can take different approaches to the same task. [Skills](/product/skills/overview) provide reusable instructions; selection supplies the immediate context.