Your second brain that watches your screen so you don't have to remember everything.
Install · How It Works · Features · CLI · API · Privacy · Download
You're deep in a research rabbit hole. Tabs everywhere. You find that one Stack Overflow answer, that perfect API doc, that Slack message with the decision your team made. Two hours later... gone. You can't find it. You can't even remember which app it was in.
Sound familiar?
ScreenMind sits quietly in your menu bar and does one thing really well: it watches what's on your screen, understands what matters, and writes it down for you.
No manual screenshots. No copy-pasting into notes. No "I'll bookmark this later" promises you never keep. It just... remembers.
ScreenMind runs a smart multi-stage pipeline that's designed to be invisible:
Event Trigger → Screen Capture → Change Detection → AX Tree / OCR → Redaction →
Skip Rules → Stealth Check → Content Dedup → AI Analysis → Storage → Export
- Triggers on meaningful events (app switch, typing pause, scroll stop, clipboard) — not fixed timers
- Captures your screen on-demand (active window, multi-display aware, screen-lock aware)
- Reads text via Accessibility APIs (10x faster) with OCR fallback (Apple Vision)
- Redacts sensitive data (credit cards, API keys, passwords) + ML-based PII detection
- Evaluates user-defined skip rules and stealth mode to filter unwanted content
- Deduplicates aggressively with 5-layer defense (hash + Jaccard + cooldown + content + buffer)
- Analyzes content with AI — app-specific prompts, vision support, context windows
- Saves to SwiftData + exports to Obsidian, JSON, Markdown, Notion, Logseq, Slack, or Webhook
- Encrypts screenshots and notes at rest with AES-256-GCM (optional)
- Indexes semantically for instant vector search and knowledge graph discovery
- Event-driven capture — triggers on app switch, typing pause, scroll stop, clipboard, not fixed timers (40-60% less CPU, 80% less storage)
- Accessibility tree extraction — reads UI text via macOS AX APIs (10x faster than OCR), falls back to Vision OCR for images/remote desktops
- 3-tier power profiles — Performance (AC), Balanced (battery >40%), Saver (battery ≤40%) with thermal override
- Screen lock detection — auto-skips capture when screen locked via CGSession polling
- Hot frame cache — in-memory cache of recent 2000 frames for instant timeline lookups
- Multi-display support (captures from display with active window)
- Active-window cropping for focused screenshots
- Timer fallback mode (5s active / 30s idle) if event-driven disabled
- Excluded apps list (skip sensitive or noisy apps)
- Manual capture with
Cmd+Opt+Shift+Cfor on-demand notes - Screenshot diffing — pixel-level comparison with connected component analysis
- 6x frame downscaling before perceptual hash comparison (30-50% faster change detection)
- Claude (Anthropic) — default, supports vision
- OpenAI (GPT-4o, GPT-4o-mini) — supports vision
- Ollama (fully offline, local models)
- Gemini (Google, via OpenAI compat) — supports vision
- Custom (any OpenAI-compatible endpoint)
- Per-provider API key, base URL, and model configuration
- One-click connection testing
- Shared prompt system (consistent quality across providers)
- Custom prompts per app — different AI behavior for Xcode vs Slack vs Chrome
- Multi-modal vision — sends screenshot to AI for visual understanding (not just OCR text)
- Context windows — AI sees last N notes for continuity across captures
- AI feedback learning — rate notes thumbs up/down, AI adapts over time
- UI element detection — identifies buttons, text fields, dialogs, menus via Apple Vision
- Screenshot diffing — highlights what changed between captures with region analysis
- Chart detection — recognizes charts/diagrams and adds specialized extraction prompts
- Structured, actionable notes from screen content
- Extracts URLs, code snippets, decisions, action items, and key data
- Smart categorization: coding, research, meetings, communication, reading, terminal
- Obsidian-compatible wiki-links and tags
- Smart tag suggestions that learn from your note history
- Meeting summaries — combines screen + audio + calendar context into action-item summaries
- Browse captures visually in a gallery or list view
- Date range filtering (today, week, month, all time)
- Category and app filtering
- Debounced full-text search across all notes
- Screenshot overlay with full-resolution viewing
Cmd+Shift+Tkeyboard shortcut
- Obsidian Markdown — daily folders, frontmatter, wiki-links, daily summaries
- JSON — one file per note for data pipelines and scripts
- Flat Markdown — simple .md without vault structure
- Notion — direct export via Notion API with database properties
- Logseq — journal format with properties syntax
- Slack — post note summaries to channels via Incoming Webhooks
- Webhook — POST note JSON to any URL (Zapier, Make.com, n8n)
- SSRF protection on webhooks (blocks private IPs)
- Multiple exporters can run simultaneously
- Content redaction — auto-strips credit cards, SSNs, API keys, passwords, emails before AI
- ML-based PII detection — NaturalLanguage NLTagger identifies names, places, organizations
- PII sensitivity levels — off, low (names), medium (+places), high (+all entities)
- Custom redaction patterns — add your own regex patterns
- Skip rules — text contains, app matches, window title matches, regex patterns
- Stealth mode — auto-pauses in sensitive apps (1Password, Bitwarden, banking, incognito windows)
- Screenshot encryption — AES-256-GCM with key in macOS Keychain
- Note encryption — AES-256-GCM for note title, summary, and details fields
- Master password — PBKDF2 key derivation (600,000 iterations) for vault protection
- Vault lock — Touch ID / password with auto-lock after configurable timeout
- Audit log — CSV trail of all captured, skipped, redacted, and exported events
- GDPR compliance — data export, retention policies, compliance reports
- All data stored locally. No cloud. No telemetry.
- Parallel OCR pipeline — configurable concurrency (default 3 workers) with backpressure
- HNSW vector index — O(log n) approximate nearest neighbor search (replaces linear scan)
- HEIF compression — 40-60% smaller screenshots vs JPEG at equivalent quality
- OCR result caching (LRU, 24h TTL)
- Screenshot storage quota enforcement (auto-cleanup at 1GB)
- CPU-aware throttling (pauses OCR if CPU >30%)
- Live CPU, RAM, and battery gauges
- Pipeline throughput (frames captured, filtered, OCR'd, notes generated)
- Notes/hour rate and session uptime
- Average OCR and AI processing times
- Skip and redaction counters
- System audio capture — hear what your Mac hears (Zoom, Meet, media) via ScreenCaptureKit
- Multi-engine STT — Apple Speech (on-device, default) with Whisper API fallback (cloud)
- Enhanced VAD — energy + zero-crossing rate with 5-frame smoothing (replaces basic RMS)
- Speaker identification — cosine similarity on audio embeddings (0.7 threshold)
- Batch transcription — buffers audio during meetings, transcribes as batch when meeting ends
- Voice memos —
Cmd+Opt+Shift+Vto record, auto-transcribes on-device - Meeting detection — Calendar integration (EventKit) detects Zoom, Teams, Meet, Slack
- Configurable language, VAD sensitivity, STT engine, memo duration
- Hybrid search — combines HNSW semantic + SQLite FTS5 keyword search with reciprocal rank fusion
- Search result caching — LRU cache (100 entries, 5min TTL) for 10x faster repeated queries
- Vector embeddings — on-device via NaturalLanguage.framework (512-dim)
- HNSW index — sub-millisecond search over 100,000+ notes
- Natural language queries — "What was I coding yesterday morning?"
- Chat with your notes — RAG-powered AI chat window with conversation history
- AI cost tracking — tracks input/output tokens and estimated spend per session
- Auto-indexes every note in both semantic and FTS5 indexes
- Auto-linking — discovers related notes via semantic similarity (>60%)
- Visual graph — force-directed layout with category color coding
- Interactive — pan, zoom, search, select nodes for details
- Project detection — auto-detects projects from Xcode/VS Code window titles
- Weekly summaries — stats: top apps, categories, tags, daily breakdown
- JavaScript plugins — sandboxed via JavaScriptCore
- 6 lifecycle hooks — onNoteCreated, onNoteSaved, onNoteExported, onAppStartup, onAppShutdown, onTimer
- Scheduled execution — interval-based scheduling (
every 30m,every 2h,every 1d) - Storage APIs —
readNote(),searchNotes(),getNoteCount()(requiresstoragepermission) - Network security — HTTPS-only (localhost HTTP allowed), 10MB response limit, 10s timeout
- MCP Server — Claude Desktop / Cursor integration on localhost:9877
- Plugin management — Settings tab with install/uninstall/configure
- Safe APIs:
log(),fetch()(with permission),getEnv()
- Notion — export notes directly to Notion databases via API
- Logseq — journal-format export with properties syntax
- Slack — post summaries to channels via Block Kit webhooks
- GitHub — create issues from notes with labels and body formatting
- Todoist — auto-extract action items (TODO/FIXME) and create tasks
- Inline editing — edit title, summary, details, tags, category after creation
- Auto-save — debounced 500ms save with visual indicator
- Calendar heatmap — month-view note density visualization, click to filter
- Focus mode integration — pauses capture during macOS DND (opt-in)
- Command palette —
Cmd+Kfor quick navigation - Tag management with add/remove
- Capture page context — URL, title, selected text, favicon, meta description
- One-click capture — send current page to ScreenMind instantly
- Right-click menu — "Send to ScreenMind" context menu item
- Connection status — shows if ScreenMind API is running
- Works with Chrome, Edge, Brave, Arc, and Chromium browsers
- If-this-then-that rules for note events
- 6 triggers — noteCreated, categoryIs, appIs, tagContains, titleContains, confidenceAbove
- 6 actions — addTag, webhook, notify, exportToFolder, slackPost, createGitHubIssue
- Rules evaluate after every note creation
- Persistent rules saved across restarts
screenmind-cli search "swift concurrency" # Full-text search
screenmind-cli list 20 # Recent notes
screenmind-cli today # Today's notes
screenmind-cli export json ~/notes.json # Export to JSON
screenmind-cli export csv # Export CSV to stdout
screenmind-cli stats # Pipeline + resource stats
screenmind-cli apps # Tracked applicationsEnable in Settings > General > Developer API.
GET /api/notes?q=search&limit=20&category=coding
GET /api/notes/today
GET /api/stats
GET /api/apps
GET /api/health
POST /api/capture (browser extension context)
For Claude Desktop / Cursor integration:
{"mcpServers": {"screenmind": {"url": "http://127.0.0.1:9877"}}}- Global keyboard shortcuts:
Cmd+Shift+N— Toggle monitoringCmd+Shift+P— Pause/ResumeCmd+Shift+S— Notes browserCmd+Shift+T— TimelineCmd+K— Command paletteCmd+Opt+Shift+C— Manual captureCmd+Opt+Shift+V— Voice memo
- Spotlight indexing for system-wide search
- Native macOS notifications
- Battery-aware capture rate
- Launch at login support
- In-app update checker (GitHub Releases API)
- Grab the latest
.dmgfrom Releases - Drag ScreenMind.app to your Applications folder
- Launch it — you'll see a brain icon in your menu bar
- Grant Screen Recording permission when prompted
- Add your API key in Settings > AI
- That's it. ScreenMind starts watching and noting automatically.
git clone https://github.com/pkmdev-sec/screenmind.git
cd screenmind
swift build -c release
# Build .app bundle + DMG + install to /Applications
./scripts/build-dmg.sh
# Or install just the CLI
cp .build/release/screenmind-cli /usr/local/bin/- Open
chrome://extensions/(or equivalent in your browser) - Enable Developer mode
- Click Load unpacked and select the
browser-extension/folder - Pin the ScreenMind icon to your toolbar
- Make sure ScreenMind app is running with the API enabled (Settings > General)
- macOS 14.0 (Sonoma) or later
- AI provider — Claude API key (default), or OpenAI, Ollama (free/offline), Gemini, or custom endpoint
- Screen Recording permission — required to capture screen content
- ~100MB RAM typical usage
ScreenMind is built as a clean Swift Package Manager project with 13 independent modules and 299 unit tests:
ScreenMindApp <- Main app (SwiftUI menu bar + windows)
|- PipelineCore <- Orchestrates event-driven capture-to-note pipeline
| |- CaptureCore <- Event monitor + ScreenCaptureKit + hot frame cache
| |- AccessibilityExtraction <- macOS AX tree walking (10x faster than OCR)
| |- ChangeDetection <- 6x-downscaled perceptual hashing + histogram diff
| |- OCRProcessing <- Apple Vision fallback + redaction + ML PII
| |- AIProcessing <- Multi-provider AI + vision + prompts + cost tracking
| |- StorageCore <- SwiftData + exporters + encryption + migration
| |- SystemIntegration <- Power profiles, lock monitor, API/MCP, vault lock
| |- AudioCore <- System audio + mic, multi-engine STT, speaker ID
| |- SemanticSearch <- Hybrid HNSW+FTS5, NL queries, RAG chat, knowledge graph
| '- PluginSystem <- JSCore engine + scheduled execution + storage APIs
'- Shared <- Constants, logging, keychain, sync engine, utilities
ScreenMindCLI <- CLI tool (search, export, stats)
Every module is actor-isolated for thread safety. The pipeline uses event-driven AsyncStream with backpressure control, 3-tier power profiles, and error boundaries with exponential retry.
All settings accessible from the menu bar > Settings:
| Tab | Settings |
|---|---|
| General | Launch at login, Obsidian vault path, data retention, disk usage, API server toggle |
| Capture | Event triggers (app switch, typing, scroll, clipboard), idle fallback, excluded apps |
| Power | Auto-switch profiles, manual override (Performance/Balanced/Saver), thermal info |
| Audio | System audio + mic toggle, STT engine (Apple/Whisper), batch mode, VAD, language |
| AI | Provider selection, API key, model, endpoint, temperature, rate limit, vision toggle, custom prompts |
| Export | Enable/disable Obsidian, JSON, Markdown, Notion, Logseq, Slack, Webhook + per-format config |
| Privacy | Content redaction, PII detection level, custom patterns, skip rules, stealth mode, encryption, vault password, audit log |
| Search | Hybrid search toggle, semantic/keyword weight balance |
| Stats | Live CPU/RAM/battery gauges, pipeline throughput, AI cost tracking |
| Plugins | Installed plugins, scheduled execution, MCP server config |
We take your screen data seriously:
- No cloud storage. Notes live in SwiftData on your Mac and optionally in your local Obsidian vault.
- No telemetry. We don't collect usage data, analytics, or crash reports.
- Your API key, your calls. Requests go directly from your Mac to the AI provider. We never see your data.
- Automatic content redaction. Credit cards, SSNs, API keys, passwords, and emails are replaced with
[REDACTED]before reaching the AI. - ML-based PII detection. Named entity recognition catches person names, places, and organizations that regex misses.
- Stealth mode. Auto-pauses capture in password managers, banking apps, and incognito browser windows.
- Screenshot + note encryption. Optional AES-256-GCM encryption with key in macOS Keychain or master password.
- Vault lock. Touch ID or password protection with auto-lock after idle timeout.
- Audit trail. Full CSV log of every capture, skip, redaction, and export action.
- GDPR compliance. Data export, configurable retention policies, and compliance reporting.
- One-click delete. Settings > Privacy > Delete All Data removes everything instantly.
- Skip rules. Define exactly what content should never be captured.
ScreenMind is open source and we'd love your help! Here's how:
- Fork the repo
- Create a feature branch (
git checkout -b feature/amazing-thing) - Commit your changes
- Push to your branch
- Open a Pull Request
Ideas for contributions:
- Windows / Linux port (Rust core extraction)
- iOS companion app for reading notes on the go
- Whisper.cpp integration for offline speech-to-text
- Firefox extension port (WebExtensions API)
- iCloud sync for multi-Mac setups
- Additional AI provider integrations
- Custom plugin marketplace
MIT License — see LICENSE for details.
Built with frustration about forgetting things, and love for the Mac.
If ScreenMind saved you from losing an important piece of context, consider giving it a star.