Use Actual
Chat
Use the chat interface once your machine and model are ready
Use Console > Chat to send text messages to a model on your computers. Chat uses your signed-in web session; it does not require a separately created inference credential.
The destination notice above the composer identifies the selected computer or cluster. Web chat sends prompts through the Actual web platform and Private Relay; the platform proxy handles the request body. Running the model on your computer does not keep a web-chat prompt on the device running your browser. The notice keeps the current request's destination while you select where to send a later request. See Private Relay for local API, offline, and multi-computer behavior.
Get ready#
Install and authorize a computer, start Actual, and download and load a model. Select the cluster you intend to use and wait for the model to report loaded.
Shared computers may be available through an organization. Access to a shared computer does not guarantee permission to change its model or settings.
Open the controls#
On a wide desktop window, move the pointer over the illustrated face at the lower left to reveal the settings panel. It contains the Cluster and Model selectors.
On mobile, swipe right to open settings or tap the left-edge Open chat settings control. Swipe left or tap Back to chat to return.
If controls are unavailable at an intermediate window width, widen the browser to use the desktop panel. If Chat is in an empty or unavailable state, first restore an online computer and loaded model.
Select a model#
The model selector reflects the chosen cluster's online computers. Installed models may be available to select without being loaded yet. Selecting one requests loading; wait for completion before sending a message.
If a computer disconnects, the available models can change. A previously saved selection is not proof that the model is ready now.
Settings#
| Setting | Initial value | Use |
|---|---|---|
| Temperature | 0.7 | Adjust sampling randomness. |
| Top K | 40 | Limit the candidate tokens considered. |
| Top P | 0.95 | Set the cumulative probability cutoff. |
| Min P | 0.05 | Set the relative probability cutoff. |
| Frequency penalty | 0 | Adjust repetition based on token frequency. |
| Repetition penalty | 1.1 | Reduce repeated tokens. |
| Presence penalty | 0 | Adjust reuse of tokens already present. |
| Seed | Empty | Optionally request repeatable sampling. |
| Stop sequences | Empty | Set strings that end generation. |
Saved settings may replace these initial values. Results and supported controls depend on the model.
Thinking shows the reasoning choices supported by the loaded model. Some models expose named effort levels and Default; Off appears only when disabling reasoning is supported. Other models use a yes/no control, which may be disabled when reasoning is unavailable.
Internet is currently disabled and set to no. Context length displays the reported model context and is read-only.
Send, stop, and copy#
Type a message and use the send control. Responses stream into the conversation. Use stop to interrupt generation and the conversation's copy controls to copy output.
Console Chat currently accepts text only and has no attachment controls. Use a compatible connected application for images or audio; see Vision & Audio.
Optional Agent mode#
Some accounts and compatible connected computers expose a Node / Agent choice. Node sends inference to the selected model. Agent can take actions such as using files and running commands on the connected computer, with activity shown in the conversation.
Agent availability depends on your account and computer's current readiness. If Agent is unavailable or sending is paused, restore the connection or switch to Node. Do not assume the Agent option is enabled for every account.
Agent tools and integrations can send data to other services. The inference destination notice does not describe every tool's network activity.
See Hermes for managed and terminal-agent setup. To use the same managed workspace in the native app, follow Hermes Desktop.
Troubleshooting#
- If you have no online computer, start and authorize one using Machine Installation.
- If the model list is empty, install a model on the selected computer.
- If loading fails, check memory availability and Logs & Telemetry.
- If a computer disconnects or the displayed model changes, wait for it to return and recheck the selection.
- If your session expires, sign in again before retrying.
- If a shared-computer action is unavailable, check permissions with its owner.
For your own application, use API Quickstart.