Agents and models

Install a model

1. Find the machine you want to use in the sidebar and open its Models action.
2. In AI Model Settings, review the available models and choose an installation suitable for that machine.
3. Wait for the download and verification to finish. The installed model list shows the machine’s available models.

Downloads and installations belong to the selected machine. Downloads can continue after you close the settings pane or disconnect your viewing client, while Dash remains running on the host. A model whose files are missing or have changed must be reinstalled before use.

The code-completion model setting is separate from the model you select in an agent conversation.

Start a conversation

Open the agent view and choose New conversation or New chat. Select an installed model from the model picker; the picker groups models by machine and shows when a machine is offline.

Before sending the first message, open Settings to choose when the agent must ask before changing files or running commands. Describe the task and the machine or folder where you want the work done.

The conversation’s settings show its system prompt and available tools. The Dash tools control determines whether the conversation can use Dash tools. You can change the model and supported thinking setting between turns.

Model machine and work machine

The model machine runs the selected installed model. The agent session has a host machine, and its tools can work with the machines and resources available to that session. The place where a model runs and the place where it reads a file or runs a command can therefore differ.

For example, you can view a conversation in Dash Web, use a model installed on a desktop computer, and ask it to inspect a folder on another available machine. Check the machine and folder shown in tool activity or an approval request. Project-scoped conversations have tighter selection limits, shown in the view.

Control changes and commands

File changes can require approval each time, be allowed, or be set to read only. Commands can require approval, be allowed, or be disabled. These permission settings are fixed when the session starts; create a new conversation to choose different permissions.

When the agent needs approval, Dash shows the tool and its arguments. Choose Approve once or Deny. Commands run with the target machine user’s permissions; choosing a project folder does not create an operating-system sandbox.

While a run is active, you can queue a follow-up, steer the run, or use Stop run. Runs can continue while their view is closed, so use the stop control when you want to stop work.

Reopen and copy conversations

Choose a conversation from the agent sidebar to reconnect to its host. Use Copy chat to copy the conversation, including thinking and tool activity. Deleting a conversation removes it on its host machine; Dash asks you to confirm.

Conversations and model files are not automatically backed up by account sync. Using a locally installed model also does not make its tools offline: commands, websites, and configured integrations can still contact other services.