Shadow
v1.0.0Your computer was paying attention the whole time.
14-modality capture. Proactive intelligence. On-device LLM. 96,000 lines of Swift and Rust. All on your Mac. Open source.
Your computer has amnesia.
You have used it for years. Done the same things thousands of times. Attended hundreds of meetings. Written thousands of messages. And every tool still treats you like a stranger. Every AI assistant asks "what do you want me to do?" as if it hasn't been sitting right there watching you do it yourself.
You are the integration layer between your apps. Slack doesn't know about your Jira tickets. GitHub doesn't know about your meeting commitments. Your calendar doesn't know you always overbook Wednesdays. You hold all the context. You do the invisible work of connecting everything together.
Shadow gives your computer a memory. Not a dumb recording, but a living, searchable, understanding memory that connects the dots across every app, every conversation, every action.
The vacuum
Rewind AI pioneered local-first screen recording with AI search on Mac. Won Product Hunt's Most Innovative Award. Raised at a $350M valuation. Then Meta acquired Limitless (Rewind's parent) in December 2025. The Mac app was killed. Screen recording disabled. EU users lost access entirely.
There is no product on the market that does what Rewind did. But Rewind was just search for your past. It recorded and retrieved. It never learned. It never acted. Shadow fills the vacuum and goes far beyond it.
14 modalities
Shadow records your Mac like a studio records a band. Each signal gets its own track. Time is the universal key. The philosophy: capture dumb, retrieve smart.
Each modality multiplies every other. A screenshot tells you what was on screen. Add keystrokes and you know what the user typed. Add the accessibility tree and you know what every element is. Add clipboard and you know what they deemed important. This is not additive, it is combinatorial.
Screen
Continuous H.265 hardware-encoded video across all displays. Sleep/wake recovery. Multi-display hot-plug.
Audio
Mic-triggered recording, no silence capture. System audio. On-device Whisper transcription with word-level timestamps.
Keystrokes
Every keystroke with passwords excluded at the CGEventTap level before anything is written to storage.
Mouse and gestures
Every click, scroll, gesture with AX element enrichment at sub-millisecond latency.
Accessibility tree
Full snapshots of every focused app, diff-aware. Captures semantic structure of every UI element.
Clipboard
Source app, destination app, content type. Knows what you deemed important enough to copy.
File changes
FSEvents on user directories. Tracks what files were created, modified, and deleted.
Git activity
Commits, branches, diffs, repo context. Knows what you produced.
Terminal commands
Commands with exit codes, working directory, output. Knows what succeeded and what failed.
Search queries
Google, Spotlight, in-app search, IDE Find in Files. Reveals your intent.
Notifications
Source app, title, whether you clicked or dismissed. Tracks your attention decisions.
Calendar
EventKit read-only integration. Knows your schedule and meeting commitments.
System context
Battery, WiFi, sleep/wake, displays, Bluetooth, volume. Environmental awareness.
Drag/drop and undo
Error recovery signal. Undo detection provides negative examples for training data.
How it works
Capture
- Under 3% CPU average
- 200-600 MB per day
- Crash-proof, loses at most 10 seconds on force quit
- Automatic sleep/wake recovery
Understand
- Episode generation with LLM summaries
- Proactive heartbeat: fast every 10 min, deep every 30 min
- Semantic search: CLIP + Tantivy full-text + timeline
- Pattern detection over weeks of behavior
Act
- 26-tool agent runtime with streaming UI
- Mimicry system learns workflows from observation
- Safety-gated replay with pre/post verification
- Built-in LoRA training on your actual clicks
On-device intelligence
Shadow runs two-tier local LLMs entirely on Apple Silicon. No cloud dependency. Your data never leaves your machine.
Qwen 7B (fast)
Quick tasks, classifications, proactive suggestions. KV-cache session reuse drops first-token latency from 14 seconds to under 1 second.
Qwen 32B (deep)
Complex reasoning, episode summaries, deep analysis. Requires 48GB+ RAM. For machines with less RAM, 7B handles everything.
MobileCLIP-S2
Image embeddings on the Neural Engine at 3-15ms per image. Powers visual search across your screen history.
ShowUI-2B + LoRA
Vision model with built-in LoRA training on your actual clicks. The agent gets better at your computer the longer you use it.
The Mimicry system
Shadow watches how you perform tasks, synthesizes replayable procedures, and executes them through a safety-gated pipeline. It learns from you, not from itself.
Safety gates
Pre-action checks, post-action verification, and undo support. The agent earns trust incrementally, not all at once.
Grounding oracle
Cascades through AX exact match, AX fuzzy, ShowUI-2B VLM, and cloud vision. Always finds the right element.
Trust ladder
Tell me things I forgot. Remind me about things coming up. Prepare things I'll need. Do things for me.
Training data
Every action is (state, action, next_state). One user generates 25,000-40,000 actions per day. Undo detection provides negative examples.
See Shadow working
Shadow in action
The interface
Search overlay. Query "ted talk" returns results across Chrome, Slack, and Shadow with visual match labels and timestamps.
Agent conversation. Shadow recalls recent activity: onboarding work, permissions, UI/UX improvements. Running Qwen2.5-7B entirely on-device.
Under the hood
Nothing else comes close
Little Bird raised $11M for the same concept. Shadow shipped first, captures more modalities, runs everything on-device, and is open source.
| Shadow | Screenpipe | Recall | Mem0 | |
|---|---|---|---|---|
| Modalities | 14 | 3-4 | 1 | 0 |
| AX tree | Full | Partial | No | No |
| Proactive AI | Yes | No | No | No |
| On-device LLM | Yes | No | No | No |
| Agent runtime | 26 tools | No | No | No |
| Open source | MIT | MIT | No | Apache |
| Price | Free | $400 | Free* | $19+/mo |
Give your computer a memory.
Requires Apple Silicon (M1+) and macOS 14+. Onboarding guides you through permissions and model downloads.