Parrot

⌘⇧E. Grammar fixed. Any app. Offline. Zero bytes sent.

---

Why I built this

Two reasons.

My girlfriend is Chinese. She writes in Italian every day — emails, messages, university assignments. She knows what she wants to say; she just needs something to catch the grammar before she sends it. Every tool she tried either required an account, sent her text to a server she didn't choose, or simply didn't work in the app she was already using.

I have dyslexia. Verb tenses, agreement, the small things that are supposed to be automatic — they're not, for me. I needed a correction layer that works everywhere I write: terminal, Xcode, a chat window, a commit message.

We needed the same thing: something quiet, offline, that works in whatever app is already open, and disappears the moment you're done. Nothing we found did all of that. So I built it.

---

Quickstart

brew install --cask thousandflowers/parrot/parrot

Or download the DMG and drag to /Applications.

First launch on macOS: if you see "unidentified developer", right-click → Open. A notarized build is in progress.

1. Launch Parrot — a small ✓ appears in your menu bar
2. Grant Accessibility access when prompted (System Settings → Privacy & Security → Accessibility)
3. Open Settings → Models → download qwen2.5-1.5b (~1 GB, fast on M1+)
4. Open Notes, type "i writed this yesterday", select it, press ⌘⇧E

Want inline autocomplete instead? Install the sibling app:

brew install --cask thousandflowers/parrot/wren

---

How it works

Parrot is a hybrid correction engine with 4 layers — fastest first, LLM only when needed:

Layer	What it does	Speed
⚡ Deterministic rules	Obvious errors (spaces, apostrophes, caps)	instant
🔤 LanguageTool	500+ grammar rules, 25+ languages, fully offline	<50ms
📖 Harper	Advanced English grammar (Rust binary, offline)	<100ms
🤖 Local LLM	Complex rewrites, fluency, style	1–3s

Each layer only runs if the previous one isn't enough. The LLM is never called to fix a missing apostrophe.

Select text in any app
        ↓
  ⌘⇧E
        ↓
Parrot reads it via the macOS Accessibility API
        ↓
Hybrid engine corrects it — 0 bytes leave your Mac
        ↓
Floating diff panel appears next to your cursor
        ↓
  Return — correction written back into the field

No clipboard. No context switch. No copy-paste.

---

Parrot vs. everything else

	Parrot	Grammarly	LanguageTool	WritingTools	TextWarden
Works in every macOS app (terminal, Xcode, Figma…)	✅	❌	❌	✅	✅
Offline, 0 bytes sent	✅	❌	partial	✅	✅
Hybrid rules + LT + Harper + LLM	✅	❌	❌	❌	❌
Accept/reject individual corrections (span diff)	✅	❌	❌	❌	❌
Per-app rules	✅	❌	❌	❌	❌
Flows (multi-step chains)	✅	❌	❌	❌	❌
Personal Knowledge Base	✅	❌	❌	❌	❌
No account, no subscription	✅	❌	❌	✅	✅
Open source	✅	❌	✅	✅	✅

---

Features

Correction modes — Grammar, Fluency, Grammar+Fluency, Translate, Writing Coach, De-Slop (strips AI filler), AI Prompt Optimizer, and fully custom prompts. Bind any mode to a keyboard shortcut.

Span-based diff panel — Accept or reject individual corrections instead of the whole suggestion. Each span shows a source badge (⚡ rule · 🔤 LanguageTool · 📖 Harper · 🤖 LLM) and an inline reason.

StyleProfiler — Detects your writing style and tone from a URL or pasted text, then adapts suggestions to match it.

Learns your style — offline. Every rejected correction is logged locally. After enough rejections it detects patterns and injects them as style notes into future prompts. No model retraining. No data leaves your Mac.

Smart Expand — Turns draft notes and short email stubs into full messages. Learns contact names from macOS Contacts.

Inline annotations — Real-time underlines as you pause typing, without triggering a shortcut. Hover to see the suggestion and apply with one click.

Flows — Chain multiple steps into a single action: grammar → simplify → translate to French. Save and trigger with a shortcut.

Knowledge Base — Store style guides, glossaries, brand voice examples. Parrot finds the most relevant entries and injects them as context into LLM prompts.

Per-app rules — Different prompt, backend, or language per app, by bundle ID.

---

Backends

Backend	Default?	Offline?	Setup
llama.cpp (bundled)	✅	✅	Download model from Settings
Apple Intelligence	auto-fallback on macOS 26	✅	Automatic
LanguageTool (bundled)	part of hybrid layer	✅	Zero setup
Harper (bundled)	part of hybrid layer	✅	Zero setup
Ollama	❌	✅	brew install ollama
OpenAI / OpenRouter	❌	❌	API key in Settings

---

Status

Current release: v0.9.3 beta

Feature	Status
Grammar / Fluency / Translate / Coach / De-Slop / AI Prompt Optimizer	✅
Span-based diff panel (per-fix accept/reject, source badges)	✅
Hybrid engine: rules + LanguageTool + Harper + LLM	✅
Inline annotations	✅
StyleProfiler + offline style learning	✅
Smart Expand + Contacts integration	✅
Knowledge Base + Flows + Story Analyzer	✅
Per-app rules, custom regex rules, presets	✅
Apple Intelligence backend (macOS 26)	✅
iCloud sync	✅
Homebrew cask	✅
Notarized release	◻︎ in progress
Mac App Store	◻︎ planned

---

Architecture

Swift/SwiftUI with actor-based shared state (zero data races by construction). AI inference runs via a bundled llama-server subprocess — no Ollama dependency, no always-on daemon, no port conflicts. Text is read and written via AXUIElement APIs directly into the focused UI element.

See docs/architecture.md for module structure, the macOS 26 constraint-loop fix, and design decisions.

---

Build from source

git clone https://github.com/thousandflowers/Parrot
cd Parrot
swift build -c release
./build-app.sh release   # produces Parrot.app + Parrot.dmg

Requirements: macOS 14+, Xcode 16 / Swift 5.10.

---

Contributing

• Bug report → open an issue
• Feature idea → open a discussion
• Easy first tasks → good first issue

See CONTRIBUTING.md for guidelines.

---

License

MIT — see LICENSE.