A high-performance, ultra-lightweight native macOS menu bar typing engine built with C++20, Qt6, and native macOS APIs (CGEventTap). Inline delivers instant typo correction, context-aware disambiguation, run-together word splitting, and inline Ghost Text predictive continuations with next-word and one-word completion modes—all running 100% on-device with zero telemetry and minimal memory footprint (<120 MB RAM active, ~0.0% idle CPU).
Warning
Windows Build Status (Currently Unstable): The native Windows port (WH_KEYBOARD_LL, SendInput) is maintained on the windows-branch and is currently unstable and under active development. For daily production typing, use the stable macOS build.
Note
Ghost Writer / Ghost Text (Beta Feature): Ghost Text predictions are currently an experimental beta feature under testing and are turned off by default. Due to varying accessibility API behavior across different hardware and macOS versions, Ghost Text will not work on all Macs. You can toggle it on in Settings > Ghost Text or via Cmd + Option + Shift + G.
- Website & Direct DMG Download
- Installation
- System Requirements & Permissions
- How to Use Inline
- Ghost Writer & One Word Mode
- Complete System Architecture & How It Works
- High-Level Architecture Diagram
- Pipeline 1: Sub-Microsecond Keystroke Interception & Security Gating
- Pipeline 2: Universal Multi-Tier Caret Localization & Blink Sync
- Pipeline 3: Dual-Tier Hybrid Prediction Pipeline
- Pipeline 4: Ghost Text Layout, Typography & Sub-Pixel Rendering
- Pipeline 5: Tier 1 Autocorrect Engine & Language Modeling
- Pipeline 6: Single-Key Undo State Machine
- Pipeline 7: Out-of-Process Neural Service & IPC
- Pipeline 8: Cross-Platform Abstraction Layer
- Key Features
- Configuration & Settings
- Building from Source
- Continuous Integration & Continuous Delivery (CI/CD)
- Packaging & Releasing
- Uninstalling
- License
Visit the official website to try the Live Interactive Simulator (running Inline's full 83,017-word and 243,342-bigram engine directly in your browser) or download the latest macOS release:
- Official Website & Architecture Walkthrough: https://laxexe.github.io/Inline-MacOS-Win32Build/
- Try the Live In-Browser Simulator: https://laxexe.github.io/Inline-MacOS-Win32Build/#demo
- Direct macOS DMG Installer (v1.5): Download
Inline-v1.5-macOS.dmg
To ensure macOS allows Inline to run smoothly without Gatekeeper quarantine warnings, install Inline directly via Terminal using Homebrew or the one-line installer.
Run this single command in Terminal:
brew install LaxExe/tap/inlineTo update Inline in the future:
brew upgrade --cask inlinecurl -fsSL https://raw.githubusercontent.com/LaxExe/Inline-MacOS-Win32Build/mac-branch/install.sh | bashThis script automatically downloads the latest release, installs it to /Applications, clears the macOS quarantine flag, and launches the app.
git clone https://github.com/LaxExe/Inline-MacOS-Win32Build.git
cd Inline-MacOS-Win32Build
git checkout mac-branch
chmod +x install_mac.sh
./install_mac.sh- Operating System: macOS 12 Monterey, macOS 13 Ventura, macOS 14 Sonoma, or macOS 15 Sequoia
- Architecture: Apple Silicon (M1/M2/M3/M4) or Intel (x86_64)
Because Inline operates system-wide to detect typing mistakes and display non-intrusive cursor suggestions in real time, macOS requires two standard permissions:
- Accessibility: Allows Inline to synthesize backspace/replacement keystrokes and inspect caret position for ghost text placement.
- Input Monitoring: Allows Inline's low-latency keyboard tap to monitor keystrokes across applications.
When launched for the first time, Inline displays a guided setup window:
- Open System Settings -> Privacy & Security.
- Select Accessibility, click
+, and add/Applications/Inline.app(enable the toggle). - Select Input Monitoring, click
+, and add/Applications/Inline.app(enable the toggle).
If permissions ever need to be reset:
tccutil reset Accessibility com.inline.mac
tccutil reset ListenEvent com.inline.macOnce launched, Inline lives discreetly in your macOS menu bar with an adaptive status indicator (|▪).
- Automatic Correction on Space: Type naturally in any application. When you press the Spacebar, Inline evaluates the completed word. If a typo is recognized, it immediately replaces it with the corrected word.
- Predictive Ghost Text on Tab: As you type, Inline shows predictive word suggestions in soft gray inline text immediately to the right of your cursor. Press Tab to insert the suggestion.
- One Word Mode: When you prefer to complete only the active word being typed (e.g. typing
thasuggestsnk), enable One Word Mode. Pressing Tab finishes only that word without chaining additional words. - Single-Key Undo via Backspace: If Inline makes a correction you did not want, simply press Backspace once. Inline immediately reverts to your exact raw typed text.
- Enter / Return Never Autocorrects: Pressing
EnterorReturn(\n) commits whatever text you entered as-is, immediately clears the internal buffer, and passes the Return key straight through. - Run-Together Word Splitting: Concatenated words (e.g.
whenIortwowords) are automatically split into separate words (when I,two words). - Sentence Capitalization: Capitalizes the first letter of sentences and quotes following punctuation (
.,!,?).
| Shortcut | Action |
|---|---|
Cmd + Option + Shift + K |
Toggle Autocorrect Engine On / Off |
Cmd + Option + Shift + G |
Toggle Ghost Text Predictions On / Off |
Tab |
Accept visible Ghost Text suggestion |
Backspace |
Revert last autocorrection or accepted suggestion (Single-Key Undo) |
Note
Beta Status & Default State: Ghost Text predictive continuations are turned off by default (ghostTextEnabled = false). It is an experimental beta feature under active testing and will not work on all Mac hardware models or macOS versions. Users who wish to test it can enable it in System Settings > Ghost Text or press Cmd + Option + Shift + G.
Inline includes an optional dual-engine prediction pipeline:
-
Layer 0 (Sub-Millisecond N-Gram & Prefix Predictor):
- Evaluates active word completions and bigram continuation phrases in
$< 100,\mu\text{s}$ . - Positioned pixel-accurately immediately adjacent to your text caret across Apple Cocoa controls, Chromium browsers, and Electron text boxes.
- Evaluates active word completions and bigram continuation phrases in
-
Layer 1 (On-Device Neural Model):
- Uses an isolated helper process running quantized
SmolLM2-135Mviallama.cpplocally on 2 Apple Silicon efficiency cores (~15 ms latency). - Zero external apps (no Ollama, no Docker, no cloud API keys).
- Automatic thermal throttling and battery protection.
- Uses an isolated helper process running quantized
-
One Word Mode:
- In Settings under Suggestion Length, select One word only (current word) or set Accept Granularity to One word.
- Ensures suggestions only finish the word currently being typed at your cursor.
Inline is architected as an ultra-low-latency, zero-telemetry native OS service. The entire engine is designed around one strict requirement: never degrade the user's typing experience. Every keystroke evaluation, caret localization probe, and layout calculation completes well within the human tactile threshold (
+====================================================================================================+
| OPERATING SYSTEM KERNEL |
+====================================================================================================+
|
Hardware Keystrokes | (Raw Scan Codes & Modifier Bits)
v
+====================================================================================================+
| PIPELINE 1: LOW-LATENCY KEYBOARD TAP |
| macOS: CGEventTap (kCGSessionEventTap) | Windows: WH_KEYBOARD_LL (Low-Level Keyboard Hook) |
+====================================================================================================+
|
| Latency <= 2 microseconds
v
+-------------------------------------------------------+
| Synthetic Loopback Gate |
| Rejects injected synthetic keystrokes (Event Source) |
+-------------------------------------------------------+
|
v
+-------------------------------------------------------+
| Security & Password Gatekeeper |
| Suspends on AXSecureTextField, PINs, or blacklists |
+-------------------------------------------------------+
|
| Active typing stream
v
+====================================================================================================+
| MAIN TRAY APPLICATION ORCHESTRATOR |
| (Lock-Free Ephemeral Word Buffer) |
+====================================================================================================+
| |
| On Word Delimiter (Space) | On Every Character Typed
v v
+-------------------------------+ +====================================================+
| PIPELINE 5: TIER 1 ENGINE | | PIPELINE 2: UNIVERSAL CARET TRACKER |
| - SymSpell 82K Frequency Dict| | - Multi-tier macOS Accessibility (AX) Ladder |
| - 243K Bigram Context Matrix | | - Chromium Selection Sync Probing (effectiveLoc) |
| - Damerau-Levenshtein Dist 2 | | - Single-line search bar vertical baseline center |
| - Run-together word splitting| | - Forward monotonicity coordinate lock |
| - Sentence auto-capitalization +====================================================+
+-------------------------------+ |
| | Exact (X, Y, LineH, Font)
v v
+-------------------------------+ +====================================================+
| PIPELINE 6: UNDO STATE | | PIPELINE 3: DUAL-TIER PREDICTIVE ENGINE |
| - Revert on single Backspace | | |
| - Restores exact raw input | | [Tier 0: Sub-Millisecond N-Gram / Bigram Graph] |
| - Zero cursor shift | | - Trie prefix lookup (< 100 microseconds) |
+-------------------------------+ | - One Word mode filter (active word ending) |
| | |
v | [Tier 1: Out-of-Process Neural Helper (IPC)] |
+-------------------------------+ | - SmolLM2-135M (Q8_0) quantized model |
| SYNTHETIC TEXT INJECTOR | | - llama.cpp inference on 2 CPU efficiency cores |
| - Atomic Backspaces | | - Prefix KV-cache reuse & token healing |
| - Replacement string emit | +====================================================+
+-------------------------------+ |
| Suggested continuation
v
+====================================================+
| PIPELINE 4: SUB-PIXEL GHOST OVERLAY |
| - Borderless click-through NSPanel |
| - Sub-pixel CoreText typography matching host app |
| - advanceLayout: < 0.2 ms type-through shift |
| - Exact vertical baseline alignment (+- 0.5 pt) |
| - Smart word-boundary leading space preservation |
| - Swallows Tab to accept continuation |
+====================================================+
-
Native Hook Mechanism:
-
macOS: Installed at
kCGSessionEventTapusingkCGHeadInsertEventTapand listening forkCGEventKeyDown. Inspection takes$\le 2,\mu\text{s}$ per key on the dedicated event dispatch loop. -
Windows: Handled via
SetWindowsHookExW(WH_KEYBOARD_LL, ...), processing low-levelKBDLLHOOKSTRUCTstructures.
-
macOS: Installed at
-
Synthetic Loopback Barrier:
- When Inline replaces a typo or inserts a prediction, it emits synthetic keystrokes. To prevent recursive feedback loops, synthetic events carry an identification signature (on macOS via
kCGEventSourceUserDataorCGEventGetIntegerValueField(kCGEventSourceStateCombinedSessionState); on Windows viaLLKHF_INJECTED). Inline discards these immediately.
- When Inline replaces a typo or inserts a prediction, it emits synthetic keystrokes. To prevent recursive feedback loops, synthetic events carry an identification signature (on macOS via
-
Security & Privacy Gating:
- Inline checks the focused UI element's accessibility attributes (
kAXRoleAttributeandkAXSubroleAttribute). - If the element is an
AXSecureTextField, password input, credit card field, or PIN box, Inline enters a dormant bypass state: no buffer is maintained, no predictions are generated, and all events pass through unmodified. - Active process bundle IDs are evaluated against a user-configurable exclusion list.
- Strict Zero-Disk Policy: Keystrokes are held exclusively in ephemeral RAM during the active word session and immediately overwritten. Zero keystroke history is ever written to disk.
- Inline checks the focused UI element's accessibility attributes (
Ghost text must sit precisely at the insertion point across native AppKit editors, Chromium browsers, Electron wrappers, and web search boxes. Inline implements a 4-tier Accessibility (AX) ladder with sub-pixel localization:
+-----------------------------------------------------------------------------------------+
| CARET LOCALIZATION DECISION LADDER |
+-----------------------------------------------------------------------------------------+
|
v
+-----------------------------------+
| Tier 1: Native AppKit Text Range |
| kAXSelectedTextRangeAttribute |
| kAXBoundsForRangeParameterized |
+-----------------------------------+
|
| If Chromium / Blink web input
v
+-----------------------------------+
| Tier 2: End-of-Text Range Ladder |
| Probe effectiveLoc = fullStr.len|
| Range(loc-1, 1) & Range(loc, 0) |
+-----------------------------------+
|
| If custom rich editor
v
+-----------------------------------+
| Tier 3: WebKit TextMarker Range |
| AXSelectedTextMarkerRange |
| Ancestor traversal (depth <= 8) |
+-----------------------------------+
|
| If AX coordinates unsupported
v
+-----------------------------------+
| Tier 4: Fallback 7c Search Center |
| Single-line baseline center |
| Font-scaled text offset |
+-----------------------------------+
-
Chromium Selection Delay Resolution (
effectiveLoc):- In Chromium and Blink web inputs (Google search, YouTube search, web
<input>/<textarea>), Blink'skAXSelectedTextRangeAttributefrequently reportslocation: 0, length: 0asynchronously while typing at the end of the input box. - Inline queries
kAXValueAttributeup front to obtainfullStr. WhenselRange.location == 0butfullStr.size() > 0, Inline probeseffectiveLoc = fullStr.size()againstkAXBoundsForRangeParameterizedAttribute. This extracts the ground-truth insertion point at the end of the typed characters, preventing suggestions from sticking at the field start and eliminating typing overlap.
- In Chromium and Blink web inputs (Google search, YouTube search, web
-
Single-Line Baseline Centering:
- In web search bars (
fieldFrame.size.height <= 72.0), if an accessibility query returns coordinates hugging the container's top border (caretRect.origin.y < fieldCenterY - 6.0), Inline automatically aligns to the centered line baseline:$$\text{baseline } Y = \text{fieldFrame}.y + \frac{\text{fieldFrame}.h - \text{caretRect}.h}{2} - 1.5,\text{pt}$$
- In web search bars (
-
Forward Monotonicity Enforcement:
- While typing on the same line,
caret.xis clamped so it never moves backwards behindminExpectedX. If an asynchronous accessibility frame reports a stale position, Inline pauses for up to 9 ms and re-queries to ensure lockstep motion with the cursor.
- While typing on the same line,
Inline couples instant deterministic lookups with deep local neural inference:
| Layer | Engine | Latency | Scope | Execution |
|---|---|---|---|---|
| Tier 0 | Unigram & Bigram Graph | Word completions, 2-word idioms | Main thread in-process | |
| Tier 1 | SmolLM2-135M (Q8_0) |
~15 ms | Full sentence continuations, context completions | Out-of-process helper (InlinePredictor) |
-
Tier 0 (Instant N-Gram & Prefix Graph):
- When a letter is typed, the active word prefix is evaluated against an in-memory prefix trie derived from 82,765 English unigrams and 243,342 bigrams.
- Returns instantaneous suggestions within microseconds without spawning threads or performing allocations.
- In One Word Mode, suggestions are clamped strictly to the active word boundary (
tha$\rightarrow$ nk), preventing sentence chaining when the user only wants word autocompletion.
-
Tier 1 (On-Device Neural Model):
- Bundles
SmolLM2-135M (Q8_0)running viallama.cpplocally on 2 Apple Silicon efficiency CPU threads. - Reuses prefix Key-Value (KV) cache entries across typing keystrokes, avoiding redundant context re-evaluation.
- Employs Token Healing (
TokenHealing): If the user's cursor is halfway through a word (e.g.unneces), token healing backs up to the last clean token boundary, generates the continuation, and strips the redundant prefix to prevent token fragmentation errors.
- Bundles
-
Safety & Filter Barrier (
SuggestionFilter):- Scrubbing blocks profanity (unless already present in user context) and strictly suppresses PII patterns (social security numbers, credit cards, telephone numbers, emails).
-
Window Architecture:
- Rendered using an
NSPanelon macOS (WS_EX_LAYERED | WS_EX_TRANSPARENTon Windows) atNSFloatingWindowLevel. - Attributes: borderless, non-activating (
NSWindowStyleMaskBorderless), transparent background, non-movable, click-through (ignoresMouseEvents = YES), and joining all virtual spaces (NSWindowCollectionBehaviorCanJoinAllSpaces | NSWindowCollectionBehaviorStationary).
- Rendered using an
-
CoreText Multi-Line Typography:
- Measures and renders text using native CoreText / DirectWrite font matching.
- Baseline vertical positioning is aligned to within
$\pm 0.5,\text{pt}$ of the host application's active line. - Dynamic proportional font size scaling:
- Search bars and tall inputs:
15.0–16.0pt(matching modern 16px web search font metrics). - Standard text inputs: clamped to a minimum of
14.0pt, strictly preventing microscopic 11pt/12pt superscript rendering.
- Search bars and tall inputs:
-
Sub-Millisecond Type-Through Shift (
advanceLayout):- As the user types letters that match the ghost suggestion,
GhostOverlay::advanceLayoutcalculates the accepted character width and translates the overlay window frame to the right on the main thread in$< 0.2,\text{ms}$ . - The suggestion moves in lockstep with the host application's blinking cursor bar without redrawing or waiting for accessibility tree synchronization.
- As the user types letters that match the ghost suggestion,
-
Smart Word-Boundary Spacing:
- If the cursor rests against a completed word (
afterWordChar == true) and the suggestion is a next-word continuation, exactly one leading space is strictly preserved (hello| there), eliminating visual and injection collisions. - If the document already ends with whitespace, leading spaces are automatically stripped to prevent accidental double spaces.
- If the cursor rests against a completed word (
-
Seamless Tab Chaining:
- In One Word Mode, accepting a word with
Tabimmediately advances the layout and reveals the next word of the sentence continuation without hiding or flickering the overlay window.
- In One Word Mode, accepting a word with
-
Dictionary & Distance Metric:
- Built on an optimized C++ implementation of
SymSpellwith precomputed delete hashes for Damerau-Levenshtein distance$\le 2$ . - Evaluates word candidates in
$< 100,\mu\text{s}$ over 82,765 unigram frequencies.
- Built on an optimized C++ implementation of
-
Context-Aware Bigram Disambiguation:
- Evaluates the preceding word context using 243,342 real-world bigram pairs.
- Resolves homophones and common confusion pairs (e.g.
therevs.their,yourvs.you're,itsvs.it's).
-
Run-Together Word Splitting:
- Employs dynamic programming to split accidentally concatenated words (
whenI$\rightarrow$ when I,twowords$\rightarrow$ two words,lookat$\rightarrow$ look at) by scoring token combinations against the frequency dictionary.
- Employs dynamic programming to split accidentally concatenated words (
-
Automatic Sentence Capitalization:
- Detects sentence starts at line beginnings and after terminal punctuation (
.,!,?), automatically converting lowercase words to uppercase initial characters, plus standalonei$\rightarrow$ I.
- Detects sentence starts at line beginnings and after terminal punctuation (
-
Case Preservation Matrix:
- Automatically detects and replicates the user's typing case pattern:
- Lowercase:
teh$\rightarrow$ the - Title Case:
Teh$\rightarrow$ The - Uppercase:
TEH$\rightarrow$ THE - Mixed / CamelCase: untouched
- Lowercase:
- Automatically detects and replicates the user's typing case pattern:
Inline features a dedicated finite state machine (UndoStateMachine) that eliminates the frustration of fighting with aggressive autocorrect:
[ Idle ]
|
| User types character
v
[ Typing ]
|
+---------------+---------------+
| |
| Delimiter (Space) | Tab (Ghost Accept)
v v
[ Autocorrected ] [ AcceptedGhost ]
| |
| Backspace immediately pressed | Backspace immediately pressed
v v
[ Revert Correction ] [ RevertGhostAccept ]
| |
+---------------+---------------+
|
| Atomic Keystroke Reversion
v
[ Raw Keystrokes Restored ]
- When an autocorrection or Tab accept occurs, the engine transitions to
AutocorrectedorAcceptedGhost, preserving the exact raw keystroke string in an ephemeral state record. - If the user's very next keystroke is Backspace, Inline intercepts the key, synthesizes the exact sequence of backspaces to remove the correction, and injects the original raw typed characters.
- Enter / Return Pass-Through: Pressing
EnterorReturn(\n) never autocorrects. It commits whatever raw text the user typed, clears the active buffer, and passes the Return key straight to the application.
To prevent neural model evaluation from ever blocking the real-time keyboard interception loop:
- Process Isolation:
Inline.appruns as the lightweight UI and hook process.InlinePredictorruns as an independent out-of-process helper process.
- IPC Protocol:
- Communicates over a private local Unix domain socket using length-prefixed binary message framing.
- Requests carry an atomic monotonic generation counter (
generation_). - If the user types a new character while an inference task is running in the background, the generation counter increments, and the helper or main thread immediately discards the obsolete response without rendering.
- Thermal & Battery Regulation (
PowerMonitor):- Continuously inspects macOS thermal state (
NSProcessInfoThermalState) and Low Power Mode. - If the system enters Low Power Mode or thermal state
Serious/Critical, the neural helper is suspended automatically, allowing Tier 0 N-grams to handle all predictions with zero battery penalty.
- Continuously inspects macOS thermal state (
The repository maintains strict architectural separation between core algorithmic logic and OS platform hooks:
Inline-macOS-Win32/
├── core/ # Platform-Agnostic Engine (C++20)
│ ├── tier1/ # SymSpell, Word Boundary, Undo State Machine
│ ├── predict/ # NgramPredictor, SuggestionFilter, TokenHealing
│ ├── ghost/ # GhostSession, GhostLayout, Layout Math
│ └── config/ # AppConfig, JSON Persistence
├── platform/
│ ├── mac/ # macOS Platform Layer (Objective-C++ / AppKit)
│ │ ├── caret/ # MacCaretTracker (AX Ladder, Blink Sync)
│ │ ├── ghost/ # GhostOverlay (NSPanel, CoreText)
│ │ ├── hook/ # MacKeyboardHook (CGEventTap)
│ │ ├── injector/ # MacTextInjector (CGEventPost)
│ │ └── power/ # PowerMonitor (IOPowerSources, Thermal)
│ └── win/ # Windows Platform Layer (Win32 API)
│ ├── caret/ # WinCaretTracker (UI Automation)
│ ├── ghost/ # WinGhostOverlay (Layered Transparent Window)
│ └── hook/ # WinKeyboardHook (WH_KEYBOARD_LL)
├── predictor_service/ # Standalone Neural Helper (llama.cpp)
└── gui/ # Native System Settings UI (Qt6)
- Branch Structure:
mac-branch: Primary macOS branch with full Apple Silicon & Intel optimizations.windows-branch: Native Windows implementation with Win32 hooks and UI Automation.gh-pages: Live interactive documentation and web demo.
Tested and verified across daily macOS workflows:
- Apple Native: Notes, Pages, Safari, Mail, TextEdit, Finder, Reminders.
- Web Browsers: Google Chrome, Vivaldi, Brave, Microsoft Edge.
- Developer & Chat Tools: Discord, Slack, Visual Studio Code, Antigravity, Notion, Obsidian.
-
Microsecond Response Time: Event tap keystroke inspection completes in
$\le 2,\mu\text{s}$ . - Context-Aware Bigram Model: Disambiguates homophones using preceding context words.
- Single-Key Undo: Immediate one-touch undo via Backspace to restore raw input.
- Enter Key Pass-Through: Return or Enter never triggers autocorrect.
-
180+ Protected Slang & Jargon Acronyms: Protects
gng,brb,ttyl,idk,ngl,tbh,repo,dev,pr,cli,sdk,api,async, etc. - Dynamic Menu Bar Contrast: Adaptive white vertical bar on dark menu bars, dark charcoal on light menu bars, with an emerald status indicator.
- Privacy & Security First: Completely offline, zero network access, zero disk logging, automatic security suspension in password fields.
- Modern macOS System Settings UI: Built to match Apple's modern System Settings with rounded card containers, responsive layouts, and light/dark theme modes.
Open Inline Settings from the menu bar to adjust your preferences:
- Autocorrect Engine: Enable or disable system-wide autocorrect.
- Ghost Text Predictions: Toggle inline cursor suggestions.
- Accept Prediction Key: Select
TaborRight Arrow. - Accept Granularity: Choose
Whole suggestionorOne word (Current word only). - Suggestion Length: Choose
One word only (current word),Short (1-2 words),Balanced (2-3 words), orLong (up to 5 words). - Overlay Contrast: Adjust between
Subtle,Standard, andHigh contrast. - Single-Key Undo: Toggle instant Backspace undo.
- Launch at Login: Automatically start Inline when your Mac boots.
- Custom Ignored Words: Add proprietary terms, names, or abbreviations that should never be modified.
- Disabled Applications: Specify bundle identifiers where Inline should stay dormant.
- Theme & Appearance: Switch between
System,Light, andDarkappearances.
- macOS 12.0 or newer
- Xcode Command Line Tools (
xcode-select --install) - CMake (>= 3.20), Ninja, and Qt 6 (
brew install cmake ninja qt6)
# 1. Configure release build
cmake -B build_release -DCMAKE_BUILD_TYPE=Release -G Ninja
# 2. Compile application bundle and prediction helper
cmake --build build_release --target Inline InlinePredictor
# 3. Ad-hoc codesign bundle
codesign --force --sign - build_release/gui/Inline.app/Contents/MacOS/InlinePredictor
codesign --force --deep --sign - -i com.inline.mac build_release/gui/Inline.app
# 4. Copy to Applications & launch
cp -R build_release/gui/Inline.app /Applications/
open /Applications/Inline.app# Run all unit tests
ctest --test-dir build_release --output-on-failure
# Or run individual test suites:
./build_release/tests/test_tier1
./build_release/tests/test_ghost_text
./build_release/tests/test_settings_binding
./build_release/tests/test_word_boundary
./build_release/tests/test_undo_state_machineInline includes an enterprise-grade automated CI/CD pipeline powered by GitHub Actions (.github/workflows/ci.yml):
lint-and-validate: Verifies dictionary files integrity and automatically enforces the zero-emoji architectural constraint across all source files.build-and-test: Compiles on macOS Apple Silicon runners with CMake and Ninja, executes the entire 19-test CTest harness, benchmarks corrector & predictor lookup latencies, and verifies Apple codesign integrity.package-and-deliver: On commits tomac-branchandmain, automatically packages the signed macOS.dmgand.zipbundles usingtools/create_dmg.shand publishes them as downloadable build artifacts with 14-day retention.release.yml: On semver Git tags (v*), automatically builds and publishes official GitHub Releases with generated release notes.
To create a distributable .dmg package with drag-and-drop installer:
./tools/create_dmg.shOutputs compressed dist/Inline-v1.5-macOS.dmg.
# Quit running instance
killall Inline InlinePredictor 2>/dev/null || true
# Remove application bundle
rm -rf /Applications/Inline.app ~/Applications/Inline.app
# Optional: Reset system permissions
tccutil reset Accessibility com.inline.mac
tccutil reset ListenEvent com.inline.mac
# Optional: Remove saved user preferences
rm -rf ~/Library/Application\ Support/InlineThis project is licensed under the MIT License. See LICENSE for details.