Settings

🌐 Language: English | 中文

Settings

Top-right ⚙ in the main window. Two-column layout: vertical tab nav on the left (Model / Appearance / Engine / Advanced / Security / Messaging / Plugins / Skills / Goal Supervisor / Mesh), independently scrolling content on the right; narrow windows fall back to horizontal tabs. A search box filters across all panels. Save / Test buttons are pinned to the bottom of the content pane — visible from any tab; the save button lights a dot when there are unsaved changes.

Model

Field	Description
Provider protocol	OpenAI-compatible / Anthropic. Determines whether requests go to /chat/completions or /v1/messages
Base URL	Endpoint root. Preset buttons: GLM Zhipu / DeepSeek / OrcaRouter / OpenAI / Anthropic
Model	Default model. e.g. glm-4.6 / claude-sonnet-4-6 / gpt-4o. Can be overridden per session
API Key	Encrypted via safeStorage (macOS Keychain / Windows DPAPI). Plaintext never goes into settings.json
Test connection	Sends a minimal request to verify. Always test before saving
Add model (v3.7.3+)	Model profile editing moved into a centered modal: profile name / key / URL / models / protocol / reasoning / pricing / balance check, with built-in test connection. Saved profiles appear in the list immediately
Zhipu balance	One-click query of the remaining balance for GLM keys

AppSettings.apiKey is encrypted via safeStorage.encryptString, stored as apiKeyEnc in userData/settings.json.

Engine / Behavior (Advanced)

Field	Description
Shell approval	always (default, modal every time) / never (auto-allow, use with caution)
Sandbox (Claude Code / Codex)	readOnly / workspaceWrite / fullAccess. Maps to --permission-mode (CC) / -s (Codex)
Plan mode	On → engine is read-only (CC goes plan mode, Codex goes read-only)
CLI engine plugins	Claude Code / Codex are plugin-gated — enable them in ⚙ → Plugins to add them to the engine dropdown
Run complex tasks in background (V3) (v3.8.0+)	On (default) → tasks V3 grades as deep submit to the JobManager; the session unlocks immediately and the result backfills when done. Off → synchronous wait
Computer Use background mode	Mouse/keyboard events are delivered to the target window in the background — your real cursor and focus never move (full support on Windows)
Window close behavior	Quit / minimize / tray

Pricing

Field	Description
Input price (per 1M tokens)	USD. Overrides the default (GLM 0.07, others 3)
Output price (per 1M tokens)	Same

priceUSD(model, tokensIn, tokensOut) (glm.ts:96) prefers your settings; falls back to built-in defaults when 0.

Interface

Field	Description
Language	English / 简体中文 / 繁體中文 / 日本語. Live switch (no restart)
Theme	dark (default) / light. Live preview (no flicker when toggling back and forth)

i18n internals: i18n.

Long-term memory

Button	Description
Export JSON	Writes to a user-chosen path. { version: 1, exportedAt: number, memories: Memory[] }
Import JSON	Accepts the above structure or a plain string[]. Dedupes by content. Returns { imported: N, skipped: N }

Good for machine migration, backup, sharing across providers. See Long-Term-Memory.

Skills (v3.7.2+)

The Skills tab aggregates every skill / command / agent from all sources — Claude Code, Codex, plugin-contributed, and the app's own built-in skills (read-only, shipped with the app; the first one, data-analysis, ports the V3 analysis discipline). Search, view, and edit user-level source files directly; plugin/built-in skills are read-only. Changes take effect immediately. See Skills.

Messaging

Configure Feishu and WeCom bot integrations. See Messaging-Bots for full details.

Field	Description
Feishu — Enabled	Toggle Feishu bot connection
Feishu — App ID / Secret	Credentials from Feishu Developer Console
WeCom — Enabled	Toggle WeCom bot connection
WeCom — Bot ID / Secret	Credentials from WeCom developer portal
Engine	Which agent engine handles IM messages
Stream reply	Send "thinking" placeholder first, then replace with answer
Default cwd	Working directory for agent tools

Per-session vs global

Global (setting)	Per-session (conv)
Provider protocol	Engine (Direct/CC/Codex)
Base URL	Model
API Key	cwd
Pricing
Approval / sandbox
Language / theme

The per-session model is an editable dropdown in the chat header's second row; supports both OpenAI-compatible and Anthropic protocols.

Encrypted storage tradeoff

CLAUDE.md notes:

API key is stored plaintext in userData/settings.json (known MVP constraint — swap for Windows Credential Manager before real distribution).

Actually, as of v1.0 it's already encrypted via safeStorage. The CLAUDE.md note is stale. Actual security: protected by macOS Keychain / Windows DPAPI; reading settings.json directly only gives you the encrypted blob.

Key source files

• src/main/settings.ts — getSettings / saveSettings / snapshot / encrypted read-write
• src/shared/types.ts:30 — AppSettings type
• src/renderer/settings.ts (embedded in app.ts) — settings view rendering