Settings

Neural Labs settings live inside the shared desktop at /workspace. Every active user receives the Settings cog in the dock. The account menu is kept small and contains only the sign-out action.

Members open Settings with Personalization, Security, Model Provider, and Plugins. Personalization controls per-user, device-local light/dark theme, custom wallpaper, desktop font size, account identity, and notification preferences. Security manages personal sign-in methods, passkeys, and verified phone numbers. Model Provider connects or pauses the personal ChatGPT account and configures private Alshival model/reasoning defaults. Plugins separates private, user-owned connections from global workspace capabilities. Administrators receive those same areas plus the control-plane areas below.

When that personal ChatGPT account is disconnected or paused, the desktop shows a clickable onboarding toast after account bootstrap. Selecting it opens or focuses Settings and switches directly to Model Provider.

The old /admin console is retired. Requests to /admin or any nested legacy path redirect active users to /workspace.

Settings areas

Personal → Security provides a private phone-number card: enter an international number, agree to a verification SMS, and submit the six-digit code. The card supports code autofill, resend cooldown, expiry/attempt errors, replacing the number without losing the existing verification, and confirmed removal. Pending verification is restored when Settings reopens. This does not enable phone login, recovery, or marketing messages.

Phone verification requires the administrator's Twilio SMS/MMS connection in Settings → Plugins. It does not opt you into proactive agent messages; those require the separate Personalization preference.

Area	Purpose
Personalization	Per-user desktop font size, account identity, notification preferences, and sign out
Security	Personal sign-in methods, passkeys, and verified phone management
Model Provider	Personal ChatGPT connection, follow-latest or pinned agent model, supported reasoning levels, and Claude availability notice
Plugins	Private plugins attached only to the member's agents and global plugins available to every workspace member
Overview	Account counts, authentication state, plugin state, runtime health, and recent audit events
Users	User approval, rejection, activation, disabling, and Admin/User role assignment
Authentication	Local and Microsoft login enablement plus Entra secret or certificate rotation
Workspace	OpenClaw health and ChatGPT/Codex device-code pairing
Audit log	The latest security-sensitive account and configuration activity
About	Live runtime versions, service state, documentation, and project links

Security model

Hiding administrator navigation is not authorization. Every /api/admin/* request independently verifies the live session, active status, and administrator role. Mutations also require an exact same-origin request and the session CSRF value in X-CSRF-Token.

The API returns display-safe identifiers and configuration summaries only. It never returns password hashes, provider subjects, client secrets, certificate private keys, OpenAI access tokens, or bearer tokens. The final active administrator cannot be demoted or deactivated, at least one web login provider must remain enabled, and local login cannot be disabled until an active administrator has linked Microsoft.

OpenAI device pairing remains workspace-owned. The control plane may start, cancel, and report the fixed OpenClaw flow through a private token-authenticated endpoint. The browser sees only the short-lived verification URL and user code; OpenClaw keeps the resulting credential in its persistent workspace volume. While pairing is active, Settings polls the control plane until the device code arrives and continues polling until authentication succeeds, fails, expires, or is cancelled; operators do not need to refresh the page between those states. After OpenClaw's login process exits successfully, the controller refreshes the local provider inventory before it classifies the result. This prevents a best-effort Gateway refresh failure or stale cached status from hiding a newly persisted OAuth profile. The CLI is not granted trusted-proxy access over container loopback.

Model Provider uses a parallel device-code controller scoped to the signed-in user's dedicated OpenClaw agent. The control plane forwards only that immutable user ID to an internal token-authenticated endpoint. It stores no OAuth token; Settings polls only safe state, URL, code, expiry, model-readiness, and pause metadata. Pause removes the user's Gateway access while retaining the OpenClaw credential, and Resume restores it. Interactive Alshival fails closed when the personal account is unavailable. The Workspace pairing remains the independent service identity for automations and background work.

The former MCP area is represented in Plugins as the locked, global Neural Labs Tools system plugin. Every active member may inspect its display-safe health, shared-agent attachment, provider readiness, and tool inventory. It cannot be edited, disconnected, or removed. Public endpoints, Entra scopes, and client registration controls remain absent because public MCP ingress is disabled.

The plugin catalog distinguishes scope from authentication. A private plugin is owned by one user, uses that user's credentials, and attaches only to that user's agents. A global plugin is installed for the workspace and is available to every member; only administrators may add or manage it. Future global plugins may use a reviewed workspace credential or require each member to make their own connection, depending on the provider.

Plugins and SMS

The shared Twilio connection is managed in Settings → Plugins, alongside Gmail and Outlook. Inbound text messages use the designated workspace native AI connection and a private conversation per member. See Messaging plugins. The legacy Plugins card remains available for existing phone-verification setup; both views use the same encrypted Twilio connection. MMS is not supported by the native conversation connector.

The add-plugin and remote MCP installation views are currently a product preview. They deliberately accept no URL or credential until the isolated credential broker, OAuth callback handling, tool review, confirmation policy, and per-agent attachment controls are implemented.

Plugin cards and API keys

Plugins opens on compact cards for Neural Labs Tools, Twilio SMS/MMS, Google Maps, KLIPY, and Pexels. Select a card for its connection details; All plugins returns to the selected scope filter. Every member can inspect status. Administrators manage the shared connections. Neural Labs Tools remains a locked system plugin, and Add plugin remains a preview for future installations.

For Google Maps, KLIPY, or Pexels, enter an API key and select Save key, then Check connection once the configuration has applied. Saved keys are encrypted and never returned to the browser. Google Maps uses one key with separate Places and Geocoding checks, so a partial setup reports which capability needs attention. Connection checks make small read-only provider requests, which may consume quota.

Saved keys take priority over deployment environment settings. Disconnect removes the saved key and disables that service, including any environment key. Use deployment configuration explicitly removes the Settings override and restores environment-based configuration. Both actions require confirmation. Changes normally apply within 15 seconds without restarting the workspace; status distinguishes applying, configured but unchecked, connected, and failed checks. During a configuration-service outage, tools retain their last confirmed configuration for at most 60 seconds, then become unavailable until it recovers.

Phone verification stays in Security; the Agent SMS/MMS updates preference stays in Personalization. Adding a phone does not opt into notifications. Open Settings windows refresh phone state after a change, including removal. See ADR 0029 for credential handling and runtime propagation.

Application boundaries

Personal model provider cards

OpenAI cards and their detail view include Refresh connection. This requests the owner's catalog with native refresh enabled and rechecks connection status; it does not reinstall software, disconnect, resume paused access, or save policy changes. Draft model/reasoning selections survive refresh. Stale cached responses are reported as failed refreshes rather than success. The catalog shows the verified Codex runtime version and last-check time. Runtime software updates remain an operator-managed, pinned-image release operation.

Model Provider opens on OpenAI and Claude cards. Select a card to open its connection settings; Back to providers returns to the overview. Claude remains unavailable until its subscription integration is supported. Connected personal accounts expose default provider, model and supported reasoning controls above the cards. The model selector offers a latest-compatible policy or an explicit pin; conversation and automation overrides are unchanged.

OpenAI's Disconnect action requires confirmation and removes the personal native auth profile, unlike Pause, which retains it. The server derives the owner from the authenticated session and requires same-origin CSRF protection. Pending login processes must stop before logout, account actions are serialized, and logout is restricted to that owner's agent and exact profile ID. Failed removal is reported without exposing native command output. Chats and model preferences are retained; workspace and Team credentials are not disconnected. Reconnection uses the existing ChatGPT device sign-in. Production accounts are not exercised by the mocked UI, API, runtime, or responsive browser tests.

The console/ bundle now owns only login, signup, and pending approval pages. The role-aware Settings application is built into workspace/desktop/ and uses the existing control-plane APIs. No additional service or port is introduced. Legacy /account requests and account-link callbacks open Security in the desktop. /workspace?settings=security opens it directly; existing ?settings=personalization links continue to open Personalization.

Nginx authenticates /workspace with the control-plane subrequest before serving the desktop. The same session cookie is then used for the same-origin Settings API calls.

Workspace messaging plugins

Use Settings → Plugins for the workspace Gmail or Outlook mailbox, and Twilio number. Open Messaging settings on that page for the designated workspace AI connection, messaging opt-ins, verification and test sends. The previous Connectors navigation entry redirects to Plugins. See Email and SMS plugins for provider setup and verification.

Desktop appearance

Personalization offers Light and Dark themes. Dark uses a generated Neural Labs wallpaper; custom wallpapers take precedence in both themes. Upload a PNG, JPEG, or WebP up to 2 MB, or choose Use Neural Labs wallpaper to restore the theme's built-in background. Image decoding and browser storage failures leave the previous wallpaper intact and display an error. Theme, font size, and wallpaper are scoped to the signed-in user in browser storage; these preferences are local to the device. Pop-out windows follow theme changes. App title bars are 32px on desktop and 40px on touch devices so window controls remain usable.

Native model pickers refresh every five minutes while visible and when the browser regains focus or connectivity. Refresh preserves the selected model and does not save a new default. Failed refreshes keep the last displayed list and report the failure. Subscription catalogs remain native, account-specific Codex catalogs; API-key accounts also retrieve their own OpenAI model inventory, filtered to GPT reasoning families suitable for agents. New inventory entries expose text input without inventing reasoning-level metadata. Existing native entries retain their reported capabilities. No workspace or shared credential is substituted for a personal connection. Native binaries remain pinned to reviewed releases.

Theme colors apply to desktop chrome, Settings, Files and file dialogs, Alshival, Projects, Skills, Automations, Deployments, Terminal controls, and preview controls. Terminal output, embedded VS Code/miniPaint, and previewed documents retain their own content themes rather than recoloring user content.