Agent

Enabling

1. For the Screph engine, prepare an LLM profile for the Agent role; for OpenHands or Codex, check the selected connection and agent engine availability.
2. Open the agent settings and enable the feature.
3. Check permissions on context, attachments, and actions before the first request.

Select the model and service in LLM settings.

New conversations use OpenHands by default. Screph and Codex are also available. Choosing another engine for an existing conversation creates a new conversation; the previous history stays with its original conversation.

How the request is processed

1. The agent receives the request text and only the context permitted by the current settings.
2. The response may contain an explanation or suggestion of a supported action.
3. In step-by-step mode, proposed changes are reviewed and confirmed. In a delegated task, the agent can continue permitted actions within the task's scope; stopping returns control to the user.
4. The result and the user's decision are saved in agent history.

Context and attachments

The request can include information about the selected item, project, current image, or an explicitly added attachment. The external LLM receives the allowed context according to the rules of the selected provider. The local model processes it on the user's computer.

Confirmation of actions

Submitting a task offers Step by step, At the agent’s discretion, and Full delegation. In the latter two modes, permission is limited to the current task and its data; the agent asks for user input when needed. Manual editing remains available.

• Reading the allowed context does not mean permission to change the project or files.
• Access mode and task mode are selected separately. With full access, the built-in Agent changes the project and settings through supported actions.
• If permission for an action is not set, the action is blocked.
• After applying, check the result and use undo, backup or project history if necessary.

A preview belongs to the displayed version of the data. After changing the project or image, or editing an object, the earlier proposal may become unavailable: a new request or an updated preview is needed. Confirmation applies to that change, not to all future actions.

Conversations and chat controls

History and new conversations open from the buttons above the chat. The + menu contains attachments, skills, and agent parameters; access and model selectors are nearby. Ctrl+F searches the displayed text. While work is in progress, Send becomes Stop.

Agent and Screph Code chats share a layout and have separate conversations, context, and permissions. Voice input fills in the request. Send and stop the task in the relevant chat. Stopping does not undo changes already applied.

Agent skills

The skills library opens from +. It supports viewing, selecting, enabling, and creating skills for the current agent engine. A selected skill adds instructions to the task; mentioning it by name helps select the relevant material.

A skill does not grant additional permissions or install hardware, models, or third-party services. An unavailable skill requires checking its scope and files. Packages for industrial or robotics tasks may contain only knowledge and examples, without a ready device connection.

History and local data

Screph stores local task and action history. OpenHands and Codex also keep conversations in their own engines; a message copy in the project does not replace that history. Before sharing a project, its saved messages and attachments can be reviewed. Deleting a local conversation does not delete data held by an external provider.

Limitations

• The agent can make mistakes; user review remains necessary.
• Quality and available features vary depending on the model selected.
• Delegation allows the selected task to continue within its scope. It does not establish that the result is correct or permit arbitrary actions outside the task; critical changes require supervision and review.