Agents & Models

Agents

Three built-in modes:

Mode	Capabilities
Chat	Chat and read-only tools
Edit	Read/write notes (default)
Create	Full capabilities, can create/modify plugins

Which mode when:

Chat Mode: read-only questions — counting, searching, summarizing, Q&A;
Edit Mode: reading and writing notes — journals, lists, editing content;
Create Mode: creating/modifying sub-plugins — panels, commands, tools, icons (see Creating Plugins in Conversation).
Switch via the agent button (upward menu) in the toolbar; disabled modes are hidden.
Custom agents: Settings → Agents → "＋ Add Custom Agent", check capability whitelists, and optionally fill a custom system prompt (persona) — written in English; when non-empty it shadows that mode's default prompt, leave empty to use the built-in default.
Fork as template: built-ins are immutable; click "Fork as template" on a built-in row to copy it into an editable custom agent (starting from its current effective persona text). The original stays untouched as fallback.
Prompt architecture: the system prompt is assembled in layers — shared identity & safety baseline (identical for all agents) + agent persona + response-language directive (generated dynamically from Obsidian's language) + dynamic context. Built-in personas are maintained as English markdown files and inlined at build time; your reply language is unaffected.
Default agent: Settings → Agents → Default agent dropdown (enabled ones only).

Providers (channels): Settings → Models. Each channel = name + Base URL (OpenAI-compatible) + API Key + model list + parameters.
Fetch models: "Fetch from endpoint" opens a searchable picker (typing a new name adds it as a candidate); if the endpoint exposes model metadata, the context window is auto-filled.
Per-model default: set any model as default — the fallback for new chats.
Per-session switching: via the model button (upward menu) in the panel toolbar.
Parameters: temperature (slider), max output tokens (0 = endpoint default; reasoning models share this quota between thinking and answer — 8192+ recommended, too small causes "thinks forever but no answer"), input context window (informational), custom headers (one Header: value per line).