Main Agent Settings
Last updated:
Main Agent settings shape the assistant used by default in chat. This includes identity, standing instructions, model choice, memory, tool access, skills, and file access.
Note
System instruction means the standing prompt an agent receives every time it runs. Put durable behavior there, not one-time requests.
Identity and instructions
Set Name and Description so the agent is recognizable in chat and settings. Use System Instruction to define its operating style, responsibilities, and constraints.
Click Change profile picture to choose a built-in picture or upload a PNG, JPEG, or WebP file up to 2 MB. Use this picture applies the selection from the gallery. The picture saves separately from the Main Agent save bar.
The instruction preview resolves available variables and shows the token count. Preview reads the saved configuration, so save instruction edits before opening it.
The editor offers these instruction variables:
{{ agent_name }}for the current agent name.{{ current_date }}for the current local date.{{ user_name }},{{ user_nickname }},{{ user_username }}, and{{ user_email }}for the current user.{{ user_timezone }}for the user's time zone.
When the agent needs the current time, allow its date/time tool instead of placing a changing time value in its standing instruction. This preserves prompt-cache reuse for local models.
Provider, memory, and compaction
Default Model Provider is used when a run does not choose a provider. Messaging channels and tasks left on Use agent default use it. The chat provider picker and a task's explicit provider selection override it. Choose Use app default to fall back to the app-wide provider.
Enable Compaction summarizes long current conversations when the context window needs room. Enable Memory lets the agent load and save memory across sessions. Enable Task List lets the agent maintain a task list during complex chat work.
Local only blocks this agent from using cloud model providers and internet tools even when global Offline mode is off. Local models, loopback endpoints, LAN endpoints, local tools, and local embedding providers can still be used.
Confidential Chat is available only for custom agents because the main agent runs Oscova's built-in memory maintenance.
Important
Agent memory requires both Enable Memory and an embedding provider in Model Providers.
Tools, skills, and files
Allowed Tools controls which ordinary callable tools the agent may use. Tool approval modes decide whether selected tools use their default behavior, require approval, auto-approve, or deny.
Tool discovery keeps large tool sets out of the prompt until needed. When it is on, pinned selected tools are visible immediately and the agent can search for the rest. Turning discovery off keeps saved pins but exposes all selected tools.
The tools table in this screen is editable. Select the tools the main agent may call, pin selected tools that should always be visible when discovery is on, and choose an approval mode for each selected tool:
- Default follows the tool's normal policy.
- Require approval asks before the call proceeds in interactive chat.
- Auto-approve allows the call without prompting.
- Deny keeps the tool selected in the list but blocks this agent from using it.
Agent Tools are other exposed agents this agent can call. Skills are imported skill packages assigned to this agent.
Workspace Folders and External Directories decide what files the agent can access. External directories need a path, label, and access mode. Labels must resolve to unique mount names.
In sandbox mode, selected workspaces appear to the agent under /workspace/<folder-name>, the agent data folder appears under /data, and external directories appear under /mnt/<label>.
Selecting Library grants read-only access to that workspace and adds library_search plus library_index_status. Manage Library content through Settings > Files so changes trigger indexing.
Save behavior
Changes are not saved automatically. The save bar appears only after a draft changes. If you leave before saving, draft edits can be lost.
After saving, future chat turns and task runs use the new configuration. A response already in progress keeps the configuration it started with.
If saving reports a provider or workspace mount conflict, choose a valid provider or give each external directory a unique label, then save again.