Skip to main content
AI Assistant settings control how Qualia’s AI behaves — from model selection to autonomy levels. Enable or disable web search. It works with every provider:
  • Enabled: AI can look up documentation and references
  • Disabled: AI works only with local context
Web search is useful for staying current with APIs and best practices.

​
New agents

Settings for how agents are created and configured.

​
Default model for agent-created agents

When an agent creates parallel agents, this model is used:
  • Consider using a faster, cheaper model for parallel work
  • Parallel agents often handle focused, bounded tasks

​
Default model for user-created agents

When you start a new chat or agent, this model is used:
  • Your primary model for direct interaction
  • Usually your most capable model

​
Model favorites

Star models in the picker to favorite them:
  • Favorites appear at the first in the model list
  • Drag to reorder favorites
  • Quick access to your most-used models

​
New agents require user confirmation

Control whether agent creation needs approval:
  • Enabled: You approve each new agent before it starts
  • Disabled: Agents start immediately
Pending agents show in the Agents sidebar when confirmation is required.

​
Research style

Default curiosity, independence, and parallelism for new agents.

​
Curiosity

How much the agent explores beyond your immediate request:

​
Independence

How much the agent assumes versus asks:

​
Parallelism

How much the agent splits work across additional agents:

​
Autonomous Mode

A separate toggle that supersedes curiosity, independence, and parallelism. When on, the agent runs fully autonomously — posting to Slack instead of asking.
Autonomous Mode requires Slack to be configured. See Autonomous Mode for details.

​
Per-session overrides

These settings are defaults. You can adjust curiosity, independence, and parallelism for individual chat sessions without changing the global settings.

​
Rules

Persistent instructions that guide AI behavior. See Rules for full documentation. In Settings, you can:
  • Create rules: Add new rule files
  • Edit rules: Modify existing rules
  • Import rules: From Cursor or Claude Code
  • Reveal folder: Open rules directory in file manager

​
Skills

Reusable workflows the AI can invoke. See Skills for full documentation. In Settings, you can:
  • Create skills: Add new skill files
  • Edit skills: Modify existing skills
  • Toggle auto-invoke: Control whether agents can trigger skills automatically
  • Import skills: From Cursor or Claude Code

​
Bash command allowlist

Commands you’ve approved with “Run & allow” appear here:
  • Approved commands run without confirmation in future
  • Remove commands to require confirmation again
  • Platform-specific (macOS, Linux, Windows): on Windows, PowerShell commands are parsed and allowlisted just like POSIX shell commands elsewhere. Prefer PowerShell-shaped commands such as Get-ChildItem -Recurse, Get-Content, Remove-Item, and rg.
  • Some commands always ask for confirmation and can never be added, because approving one would approve everything it can run: shells and script hosts (bash, powershell, pwsh, cmd), wrappers (sudo, env, wsl), and process launchers (Start-Process). Aliases are resolved first, so start is treated as Start-Process. This includes running a script through a host, so powershell -File build.ps1 asks every time — the host can run any script you point it at, and its switches can be abbreviated, so there is no way to approve one form without approving the rest.
  • Interpreters (python, node, perl, ruby) are the one exception: they can be added, but their inline-script switch stays blocked, so python script.py runs without confirmation while python -c "..." still asks.