Complete reference for every setting in ADE — from AI provider management and budget caps to notifications, permissions, and developer tooling.
**Screenshot: The Settings page with the left navigation sidebar visible, showing all section categories listed (General, AI Routing, Context, Budget & Usage, AI Providers, Integrations, Computer Use, CTO, Automations, Permissions, Notifications, Developer). The “General” section is active and its content is shown on the right.
Sidebar: Click the gear icon (⚙) at the bottom of the left sidebar
Menu bar: ADE → Preferences
Settings are organized into twelve sections in the left navigation panel. Changes in most sections take effect immediately — no restart required. Exceptions are noted in the relevant section below.
ADE’s default model applies to all agents unless you set role-specific overrides. Role overrides let you optimize cost and quality by matching model capability to task complexity.
Role
Description
Recommended Model
Orchestrator
Mission planning and worker coordination
claude-opus-4-6
Worker
Mission execution tasks
claude-sonnet-4-6
Validator
Mission validation and review
claude-sonnet-4-6
Automation
Event-driven automation runs
claude-haiku-4-5
Chat
Interactive chat sessions
claude-sonnet-4-6
CTO
Project-level memory and delegation
claude-opus-4-6
Routing automation runs to claude-haiku-4-5 while keeping claude-opus-4-6 for mission planning is the single most impactful cost optimization for users running frequent automations. Most PR-review and routine triage tasks do not require Opus-level reasoning.
Settings for context doc generation and refresh behavior. Context docs (PRD, architecture overview) are synthesized from your codebase and injected into agent sessions as project-level context.
Setting
Description
Default
Generation model
AI model used to generate context docs
claude-sonnet-4-6
Reasoning effort
Depth of analysis during generation
Default (model decides)
Auto-refresh events
Which project events trigger automatic doc regeneration
Track and control AI spending across all agents and providers.
**Screenshot: The Budget & Usage settings panel showing: a “Current Month” spending summary card (34.20of100.00 monthly cap), a stacked bar chart breaking down spend by provider (Anthropic, OpenAI), and below it a table showing spend by lane for the last 7 days.
Maximum USD across all workers in a single mission
$10.00
Per-automation cap
Maximum USD per automation trigger
$0.50
Monthly total cap
Hard ceiling across all providers, agents, and sessions
$100.00
The monthly cap is a hard stop. Once reached, all agent activity (chat, missions, automations, CTO heartbeats) pauses until either the cap is raised in Settings → Budget or the calendar month rolls over. ADE sends a notification when you reach 80% and again at 100% of the monthly cap.
Manage API keys for all configured providers. This section is a GUI layer over local.secret.yaml. Changes saved here write directly to that file.For full provider setup instructions, see AI Providers.
Action
How
Add API key
Click Add Key next to the provider, enter the key value
Rotate key
Click Rotate — enter the new key. ADE swaps the key in memory immediately.
Revoke key
Click Revoke — ADE removes the key from local.secret.yaml and memory. Agents using that provider will fail until a new key is added.
Test connection
Click Test Connection — ADE sends a minimal request and reports success or the error message
Set default model
Dropdown per provider — writes to ade.yaml (requires SHA re-approval if other fields changed)
Configure which computer-use backend ADE uses for agent-driven browser and desktop automation.
Setting
Description
Default
Backend
ghost-os, agent-browser, or ade-local
ade-local
Connection URL
URL for Ghost OS or remote backends
—
Auto-connect on launch
Connect to the backend when ADE starts
On
Default proof requirements
Which action types require screenshot proof
Screenshots for all write actions
Proof storage location
Where captured screenshots and traces are saved
.ade/artifacts/computer-use/
The capability matrix at the bottom of this section shows which features each backend supports (screenshots, video recording, DOM access, file download interception, etc.).For full computer use documentation, see Computer Use.
Settings for the always-on CTO agent that monitors your project and manages workers.
Setting
Description
Default
CTO enabled
Enable or disable the CTO agent entirely
On
Heartbeat interval
How often the CTO runs its check-in loop
15 minutes
Default worker budget
USD budget allocated to each worker the CTO spawns
$2.00
Memory retention
How long to retain project memory entries before summarizing
30 days
Linear sync
bidirectional, read-only, or disabled
bidirectional
Auto-delegation
Allow CTO to spawn workers without asking for approval
Off
Auto-delegation threshold
Only auto-delegate tasks with estimated cost below this value
$1.00
The heartbeat interval controls how frequently the CTO agent wakes up to review open issues, pending missions, and project health. A shorter interval (e.g., 1 minute) gives more responsive CTO behavior but increases AI spend. For most projects, 15 minutes is the right balance.
Settings for debugging, diagnostics, and low-level access.
**Screenshot: The Developer settings panel showing: a “Log Level” dropdown set to “info”, a “Log file location” path field with a folder-open button, an “Open DevTools” button, a “MCP Debug Mode” toggle (off), and a “Database Inspector” button that opens a read-only SQLite browser.
Setting
Description
Default
Log level
error, warn, info, debug, verbose
info
Log file location
Where ADE writes its log files
~/Library/Logs/ADE/
Open DevTools
Opens Electron DevTools for the renderer process
—
MCP debug mode
Enables verbose JSON-RPC logging for all MCP calls
Off
Database inspector
Opens a read-only browser of ADE’s SQLite database
—
Reset to defaults
Resets all settings to factory defaults (does not affect local.secret.yaml)
—
verbose log level logs every MCP tool call including parameters. Do not use verbose logging in production environments — it may produce large log files and can expose non-secret parameters in plaintext. API keys are always redacted regardless of log level.
