Settings Reference¶

Gear icon → Inline Agents.

Plugin-wide¶

Default Agent, Auto-approve, and Allow JS expressions settings

Setting	Default
Default Agent	Claude Code	Which CLI agent note buttons run when they don't specify their own `agent:` line.
Auto-approve by default	Off	When on, buttons run with permission checks bypassed unless a button sets `autoApprove: false`. When off, the agent asks before each tool use, right in the terminal, unless a button sets `autoApprove: true`.
Allow JavaScript expressions in prompts	Off	Gates `{{= expression }}` segments in `prompt:` templates. Plain `{{ file.basename }}`-style lookups always work regardless of this setting — see Context and Templating.

Per provider (Claude Code / OpenCode)¶

Claude Code settings: binary path, model mappings, connection status

Setting
Connection status dot	green/red next to the provider name	Whether the configured binary actually runs (`<bin> --version` exits `0`) — not an auth check. Refreshed on Settings open and via the ↻ button.
Binary path	text field	Absolute path to the executable, or a bare command name to rely on `PATH`. Falls back through known install locations and finally `PATH` lookup if the configured path doesn't exist — see Claude Code and OpenCode.
Model mappings	dynamic list	Name → model string pairs a button's `model:` line resolves against — see Model Mappings.

OpenCode's model dropdown is a live list from opencode models:

OpenCode's live model dropdown

Agent runtimes¶

A second way to run a button — hand its prompt off to an HTTP agent runtime (OpenClaw, Hermes Agent, any OpenAI-compatible endpoint) instead of the inline terminal. Full detail in Agent Runtimes; summary here:

Setting
Default target	dropdown, `Terminal` plus every configured runtime name	What a button without a `target:` line uses, and what an unknown `target:` name falls back to.
(toggle next to it)	on by default	The default for a button's `open:` line when it's missing or malformed.
Per runtime: Name, Kind (OpenClaw hook / OpenAI-compatible chat), Base URL, Agent ID or model, Token required, Send session key as `x-openclaw-session-key` (chat kind only), Send session id as `X-Hermes-Session-Id` (chat kind only, default off), Token (masked, per-device secret storage), Thread-URL template (placeholders `{agentId}`, `{sessionKey}`, `{sessionKeyPath}`, `{sessionId}`)	—	Add, edit or remove a runtime (no reordering — remove and re-add to change order). Two runtimes can't share a name.
Test connection	button, per runtime	Checks reachability and the token without sending a prompt.

Where each setting lives in the source¶

For contributors: the whole tab is AgentConsoleSettingTab in src/settings.ts, using the classic imperative display() API rather than the declarative getSettingDefinitions() one — the model-mapping list (dynamically added/removed rows, a custom dropdown with a per-option copy button, live status dots) needs DOM control the declarative control types can't express. One trade-off: this tab doesn't appear in Obsidian's own in-app settings search.