Settings Reference

Open Settings with ⌘,. It has four tabs — AI, Terminal, Security, and Privacy — plus a footer with Reset to Defaults and Apply.

Apply vs. immediate: most AI and Terminal changes take effect when you press Apply. A few controls apply immediately because they change the live safety gate or session behavior: host profiles, policy packs, the "restore last session" toggle, and the cloud-context picker. Each section below notes which is which.

AI tab

Controls your active AI connection and model. See AI Providers and Models for full per-provider setup.

The tab has two columns, each scrolling on its own: the list of connections on the left, and the selected connection's settings on the right.

• AI Provider — the left column: a list of connection tiles, one per row, grouped (managed local, native cloud, OpenAI-compatible, custom, CLI subscriptions). Picking a tile changes the right column only. Selecting a Pro provider while Free prompts an upgrade.
• New custom server… — under the list, always in view: adds your own OpenAI-compatible server.
• Set Up AI… — under the list too: re-runs the guided setup wizard.
• Escalation model — the stronger connection a stalled /fix can be handed to, and its model. For Claude Code, choose Sonnet or Opus; Codex uses its own default model.
• System Prompt — a multi-line editor with Reset to Default.
• Connection check — under the selected connection's settings: what AITerm knows about the applied connection, and Check, which sends one short message (“Reply with the single word OK.”) to it and shows whether it answered — a green check mark when it worked, an orange mark and the reason when it didn't, with the fix beside it (Sign in / switch account, Set Up AI…, Review…). Cancel stops a check that is still running. While the form has unapplied changes, a typed key included, Check waits for Apply, which checks on its own. The same mark sits on the status bar's connection chip, whose menu has Check connection.
• Named servers — a saved server's address is read-only; Change address… opens the sheet. When a saved server's address differs from the one in use, the notice under its key field offers Use saved address — one click that stages the saved address (the same as clicking the server's tile); enter the key and click Apply (Add Connection for an Azure server).
• Server status — for remote/custom providers: Health (with Refresh) and the Endpoint URL.

Per-provider configuration appears inline for the selected provider:

• Local AI Engine panel for Ollama — Health (with Refresh), Endpoint (with Copy), Version, and Models path (with Reveal in Finder). It also has Reinstall Engine…, which keeps your installed models.
• Azure OpenAI — Name, Resource, Deployment, API version (default 2024-10-21), API key, and an Add Connection button that stays disabled until resource, deployment, and key are filled.
• Amazon Bedrock — Name, Region, a Credentials picker ("Use an AWS profile" / "Paste access keys"), the matching credential fields, and a Model / profile id field.
• Managed MLX — current Model plus Choose model….
• Claude Code / Codex CLI — nothing to configure, but the CLI must be signed in. Check confirms that before the first request, and a Sign in / switch account action opens a tab and runs the CLI's own login.
• OpenAI / Anthropic / OpenAI-compatible — API key, an optional preset and base URL, Save key & load models, and a Model picker.

Apply-gated AI changes take effect on Apply. Provider setup sheets save immediately through their own Save button.

Terminal tab

• Default Shell — /bin/zsh, /bin/bash, /bin/sh, or /bin/fish.
• Theme — Dark, Light, Solarized Dark, Solarized Light, or Ocean.
• Font Size — a slider from 8 to 24 pt. This also responds live to the ⌘+ / ⌘− / ⌘0 View shortcuts.
• Restore last session on launch — reopen your tabs, splits, and directories next time. Applies immediately.
• Notifications — one honest row about long-running-command finish notifications: an Enable notifications… button if you've never been asked, a link to System Settings if they're off at the system level, or a confirmation that they're on.
• Software Update — shows the current version, a Check for updates automatically toggle (on by default: hourly background check, surfaced as a gentle "Update ready" badge), an Automatically install updates toggle (off by default: you get the badge instead), and Check for Updates…. See Updating and Uninstalling.

Each Settings tab starts with a one-line note saying which of its controls apply immediately and which wait for Apply. Drag the window's bottom edge to make it taller or shorter; it remembers the height and the tab you last used, Esc (or Close) closes it without applying staged changes, and Apply saves and closes.

Security tab

Controls the gate and its tightening layers. See Safety and Security and Remote Commands.

• Enforce Security Restrictions — the master switch for command checks.
• Start in Dry Run Mode — commands are shown but not executed by default.
• Confirm High-Risk Commands — require confirmation for high-risk commands (needs Enforce on).
• Allow Critical Commands (with confirmation) — allow the most severe Irreversible commands only after explicit confirmation (needs Enforce and Confirm High-Risk on). The help text labels this "EXTREMELY DANGEROUS" for good reason.
• Per-host SSH profiles (Pro) — map a host pattern (e.g. prod-*) to a safety floor, a pane tint, and a label. Includes an Add prod-* template button. Free users can view and delete profiles but not author them.
• Policy pack (import free · author Pro) — Import a pack… to add a floor, host rules, and deny/ask patterns; authoring, adding rules, and exporting are Pro. Deleting a pack or a rule is always allowed.

The four toggles apply on Apply; host profiles and policy-pack edits apply immediately (they change the live gate).

Privacy tab

See Privacy and Data Handling.

• Where your data goes — three information rows: on-device by default, your keys stay in the Keychain, and a clear on-device-vs-cloud badge for every request.
• Redacted before any cloud request — the list of secret categories scrubbed before a cloud send.
• Session context sent to cloud models — a picker: Recent commands + output, Recent commands only (default), or Nothing. This caps only what a cloud model receives; local models always get full context. Applies immediately.

Footer

• Reset to Defaults — resets settings. If host profiles or a policy pack are active, it asks first (they're removed immediately; other fields still wait for Apply).
• Apply — saves the Apply-gated changes. A brief "Applied" confirmation appears.

Related pages: AI Providers and Models, Safety and Security, Privacy and Data Handling.