Moro User Guide

Getting started

Moro is a macOS menu-bar app. After you grant the required system permissions, its shortcuts can be used for dictation, translation, rewriting, and questions in most apps with standard text fields. Some apps expose limited system interfaces, so certain features may be unavailable or use a compatibility mode.

Requirements

• macOS 15 (Sequoia) or later — the built-in local speech recognition needs it.
• Apple Silicon (M-series) — local speech and local LLM inference are arm64-optimized; Intel Macs can't run local inference.
• Speech recognition and text generation are configured separately. You can stay fully offline with local models, or use your own key for cloud models.

First-time setup

Onboarding recommends a path from the app language. Chinese, Japanese and Korean start with the quick cloud setup; English starts with on-device speech. You can switch paths at any time. This choice sets up the first speech recognizer only and does not lock the model used by every feature.

1. Complete the two-service cloud check, or wait for the local ASR download to finish.
2. Speak the sample line into your microphone. The first try counts only after real audio produces a transcript.
3. Grant the four core permissions: Microphone, Accessibility, Input Monitoring and Full Disk Access.
4. After entering Moro, test voice write-back with ⌥⌘ in a standard macOS text field, then open Quick Chat with ⌥C.

Permissions on first run

Onboarding walks you through four core permissions (Microphone, Accessibility, Input Monitoring, Full Disk Access) — each tied to a specific capability and explained as you go; all four make the experience complete. Anything else is requested only when you use the matching feature:

• Microphone: required for voice input.
• Accessibility: to power the global shortcuts, read the context at your cursor and write results back into the field. This is Moro's core channel — it uses the system Accessibility API, not screenshots.
• Input Monitoring: keeps the global shortcuts reliable in edge cases with certain keyboards/IMEs (like Space×3 under some Chinese IMEs).
• Full Disk Access: powers the local "find file" search — names, file contents and recently-opened files are all matched against your Mac's own Spotlight index, nothing leaves the machine; without it, files in Documents/Desktop/Downloads can't be found.
• Screen Recording (optional, off by default): used only if you opt into the experimental reply-draft feature inside closed apps like WeChat. Everyday input never touches it.
• Automation (Apple Events) (optional): asked for only when you bring Finder-selected files into Quick Chat, Smart Voice or Smart Text — Moro requests permission to control Finder at that moment, not on first launch.

All shortcuts

The table below lists the default shortcuts. All shortcuts except ⇧⇧ for autocomplete can be changed in Settings.

The voice key supports hold-to-talk out of the box: hold the modifier to speak and release to stop, while a quick tap still toggles start / stop. You can turn it off in Settings.

Voice input

Press ⌥⌘ and speak. Moro transcribes, polishes, and writes the finished text at your cursor — fillers removed, punctuation and numbers fixed, usually ready to send as-is.

⌥ ⌘

✦ Voice Listening…

The dictation capsule floats at the bottom center of your screen (draggable, remembers its spot per display) without stealing focus: icon · ✦intent · status · ember. It tucks away after writing the finished text back.

• Adjustable polish strength: from near-verbatim to fully restructured, per your taste.
• Language recognition: supports English, Chinese, and mixed input. Accuracy varies by speech model, accent, and recording conditions.
• Local or cloud transcription: choose a local or cloud speech service in Settings. With a cloud service, audio is sent to that provider for processing.
• Recordings are kept locally only for re-transcription; retention is configurable or can be turned off.
• Optional sound cues: go to Voice → Recording and choose Off or one of ten themes. Each start cue has a matching completion answer.

Smart actions

Press ⌥Space to speak, or double-tap ⌘ to type. Depending on the current text field and selection, Moro can perform these actions:

• Write / dictate: turn speech into text, written into the current field.
• Act on selection: translate, rewrite, polish, summarize, reply, explain — on the text you selected.
• Ask: send the question to Quick Chat, answered in an overlay.
• Note: jot one down, recall an old one, or check a to-do off.
• System actions (Beta): open apps, find files, and open settings or websites.

Use a clear action phrase, such as “translate into English,” “open Calendar,” or “find last week's contract.” If the result is not what you expected, press ⌥Z to undo the latest write-back.

Say this, get that

A verb at the start of the sentence triggers most reliably ("translate into English" beats "this text, I'd like to translate it"); polite prefixes (please / could you) are stripped automatically.

Common tasks, step by step

Typing autocomplete Beta · off by default

When speaking is awkward — a meeting, late at night, or you haven't found the words yet — just type. Pause, and Moro predicts your next line in light gray: a small card floats above the input, stitching your text together with the prediction (right at the end of a line it can also appear inline after the cursor); press Tab to accept. Voice is for "say it once you know it"; autocomplete is for "draft a sendable line while you're still finding the words." They complement each other.

Turning it on & using it

• Off by default. Enable it under Autocomplete in Settings (it needs Accessibility and runs an always-on key watcher, so it's opt-in).
• Trigger: it predicts after a typing pause of about 0.35s; the cadence (Instant / Standard / Steady / Relaxed) is adjustable.
• Tab accepts all · ⌥Tab word-by-word · Esc dismisses. Keep typing and it refreshes in place; arrow keys, a click or switching apps hides it.
• Mid-sentence works too: it reads text on both sides and fills only the middle, leaving what follows untouched.
• In closed apps like WeChat where the caret position isn't available, double-tap Shift to summon it — this closed-app fallback is experimental and off by default; first enable it under Autocomplete › Closed apps like WeChat (experimental), or grant Screen Recording.

Availability

Autocomplete uses Accessibility to read text near the current field and does not read the clipboard. It does not run in password fields or restricted apps. Some apps do not expose the caret position or text context; autocomplete may be unavailable there, although you can try ⇧⇧ to summon it manually.

Translate, rewrite, polish, enhance

Select some text (Translate doesn't even need a selection), press the matching shortcut, and the result replaces in place or floats beside it. ⌥Z undoes the latest write-back (within about 2 minutes); where an app can't be reverted in place, your original is copied to the clipboard instead.

Quick Chat & screenshot ask

⌥C to ask in a flash — attach your selection, files multi-selected in Finder, or an image; use the answer right away. Esc collapses it to a floating pill that stays where the panel was — drag it to the left or right screen edge to dock it — so it never interrupts your work.

• Multi-model: pick a different model per chat; built-in code highlighting and Markdown rendering.
• Web search: connect Bocha, Zhipu, Tavily, Serper, Jina, or key-free DuckDuckGo; default is Bocha in China, Tavily elsewhere.
• Screenshot ask ⌥A: region-select part of the screen and ask about it by voice.
• Private Chat ⌥V: burns on close, no history.

Notes & to-dos

Catch a thought: press ⌥D and just say it. Moro files it into a card — idea, note, meeting, to-do or plan, sorted for you. Do it all by voice: create, complete, reschedule, search.

Models & routing

Moro doesn't host models or resell tokens. Cloud is bring-your-own-key, with presets for OpenAI, Anthropic, Gemini, Qwen, DeepSeek, Kimi, GLM, MiniMax, MiMo, StepFun, Mistral, SiliconFlow, and more, plus OpenAI / Anthropic-compatible endpoints and direct OpenAI Codex subscription sign-in. It also connects to local Ollama / LM Studio. The local inference engine is built in—no Ollama required; model sizes vary, and downloaded models can run fully offline.

Route per feature: polish, rewrite, translate, chat, voice-polish and autocomplete can each point to a specific cloud or local model. Send everyday short text to a cheap fast model, save a strong model for heavy work, keep private content local — mix freely.

ASR and LLM do different jobs

Choose by region and language

ASR turns speech into text. The LLM handles polishing, translation, and chat. You can choose them separately.

Privacy & security

• Local-only in one tap: ASR, LLM, and web search stay offline. Saved cloud settings remain available but are not used until Local-only is turned off.
• Cloud models: when you use a cloud model or speech service, relevant content is sent to the provider you configured and is subject to that provider's privacy and retention policies. Review those terms before use.
• Password fields always skipped: secure fields are detected and never read; password managers and other sensitive apps are forced local or skipped entirely.
• Local history is encrypted: sensitive text (transcripts, chats, learned samples) is stored with field-level AES-GCM on your Mac, key in the Keychain, deletable anytime. Optional voice recordings kept for re-transcribe are stored unencrypted—turn recording retention off, or pick a short auto-cleanup window, if you'd rather not keep them.
• Screen content: normal text access uses Accessibility. The screen-based compatibility mode for closed apps is off by default and must be enabled manually. When you use ⌥A, the selected image is processed by the active chat model.

Undo

Undo write-back ⌥Z: what was just written can be reverted, in the same app and shortly after — the last 5 actions are held in memory (never persisted). If the field or app changed and it can't revert in place, then for edits to a selection (translate / rewrite / polish) Moro copies the original to your clipboard and tells you; for plain voice dictation, if it can't safely delete what it inserted, it tells you it couldn't undo rather than copying. You can always edit by hand, too.

Pricing & license

• One-time purchase, not a subscription. License scope and upgrade terms will be confirmed at launch.
• Pricing will be announced later. Apply for private beta access before launch.
• 7-day free trial.
• One license per machine, transferable after you deactivate a seat. After an online check it works offline for a while (about two weeks including grace); transient network or server hiccups never lock a paying user out.
• Cloud costs are separate and yours to control: bring your own key and pay per use, or just run the free local models.

FAQ