Skip to main content

Skill Reference

Skills are organized by category and tagged with their requirements. This page provides detailed technical specifications for every skill available in the Aegis AI marketplace.

Detection Skills

YOLO 2026 Object Detection

Real-time object detection using the latest YOLO architecture. Detects 80+ COCO object classes with colored bounding boxes, class labels, and confidence scores rendered directly on camera feeds.

SpecificationValue
CategoryDetection
PlatformsmacOS (Apple Silicon + Intel), Windows (x64), Linux (x64 + ARM64)
RequirementsPython 3.9+
UnlocksDetection Studio view, detection overlay on Monitor tiles

Model sizes: Choose the right tradeoff for your hardware:

SizeSpeedAccuracyBest For
NanoFastestGoodCPU-only systems, Raspberry Pi, resource-constrained setups
SmallFastBetterGeneral use, balanced speed and accuracy
MediumModerateHighGPU users who want detailed detections
LargeSlowestHighestMaximum accuracy, dedicated GPU recommended

Adjustable settings:

  • Confidence threshold (0–1) — minimum detection confidence to display a bounding box. Higher values reduce false positives.
  • Processing interval — how often frames are analyzed (presets from 0.2 to 15 fps)

80+ COCO classes include: Person, bicycle, car, motorcycle, bus, truck, dog, cat, bird, backpack, umbrella, handbag, suitcase, bottle, cup, fork, knife, laptop, phone, book, clock, and many more.


Privacy / Transformation Skills

Depth Estimation (Privacy)

Privacy-first depth map transforms using Depth Anything v2. Converts camera feeds into colorized depth maps where near objects appear warm and far objects appear cool. In "depth only" mode, the scene is fully anonymized — you can monitor spatial activity without revealing anyone's identity.

SpecificationValue
CategoryPrivacy
PlatformsmacOS (Apple Silicon + Intel), Windows (x64), Linux (x64 + ARM64)
RequirementsPython 3.9+
Unlocks3D Depth Vision view

Display modes:

ModeDescription
Depth OnlyFull anonymization — shows only the depth map, no identifiable visual information
OverlayDepth map blended over the original image at adjustable opacity (0.0 to 1.0)
Side by SideOriginal camera feed and depth map shown adjacent for comparison

Colormaps: Inferno, Viridis, Plasma, Magma, Jet, Turbo, Hot, Cool — each provides a different visual style for depth visualization.

Hardware acceleration:

  • macOS (Apple Silicon): Uses CoreML on the Neural Engine — extremely fast, leaves GPU free
  • Windows/Linux (NVIDIA): Uses CUDA for GPU acceleration
  • CPU fallback: Works on any system, but slower

Depth models available:

  • depth-anything-v2-small — fastest, smallest memory footprint
  • depth-anything-v2-base — balanced quality and speed
  • depth-anything-v2-large — highest quality depth estimation

Segmentation Skills

SAM2 Segmentation

Interactive segmentation using Meta's Segment Anything 2. Click any point on a camera frame, or draw a bounding box, to get pixel-perfect segmentation masks. SAM2 excels at separating objects from backgrounds with surgical precision.

SpecificationValue
CategorySegmentation
PlatformsmacOS (Apple Silicon + Intel), Windows (x64), Linux (x64 + ARM64)
RequirementsPython 3.9+
UnlocksSegmentation Studio view

Capabilities:

  • Point prompts — click a single point on an object to segment it
  • Box prompts — draw a rectangle around an area to segment everything inside
  • Video tracking — select an object once and SAM2 tracks it across consecutive frames
  • Multi-object — segment multiple objects simultaneously with different masks

Annotation Skills

Annotation Data

Dataset annotation management for building custom training datasets from your camera footage.

SpecificationValue
CategoryAnnotation
PlatformsmacOS (Apple Silicon + Intel), Windows (x64), Linux (x64 + ARM64)
RequirementsPython 3.9+
UnlocksAnnotation Studio view

Capabilities:

  • Import camera frames for annotation
  • COCO-format label management
  • Organize annotations into sequences
  • Export datasets for model training
  • Upload to Kaggle for sharing and collaboration

Training Skills

Model Training

Agent-driven YOLO fine-tuning workflow. Train custom detection models on your own camera data so the detector learns objects specific to your environment.

SpecificationValue
CategoryTraining
PlatformsmacOS (Apple Silicon + Intel), Windows (x64), Linux (x64 + ARM64)
RequirementsPython 3.9+

Export targets:

  • TensorRT — optimized for NVIDIA GPUs
  • CoreML — optimized for Apple Silicon
  • OpenVINO — optimized for Intel hardware
  • ONNX — universal cross-platform format

Workflow:

  1. Annotate your camera data using the Annotation Studio
  2. The agent guides you through training configuration
  3. Training runs on your local hardware with progress reporting
  4. Trained models are automatically exported in the optimal format for your GPU
  5. Deploy the trained model as a custom detection skill

Analysis Skills

Home Security AI Benchmark

Evaluation suite for testing LLM and VLM performance on real home security tasks.

SpecificationValue
CategoryAnalysis
PlatformsmacOS (Apple Silicon + Intel), Windows (x64), Linux (x64 + ARM64)
RequirementsNode.js 18+

Tests include:

  • Deduplication accuracy — can the model correctly identify duplicate events?
  • Scene classification — how accurately does the model categorize activity types?
  • Tool usage — does the model correctly invoke tools when needed?
  • Description quality — how detailed and accurate are scene descriptions?

Integration Skills

Camera Claw

Security camera integration for OpenClaw robotic arm systems. Sandbox, record, and monitor OpenClaw activity with full AI observation.

SpecificationValue
CategoryIntegrations
PlatformsmacOS (Apple Silicon + Intel), Windows (x64), Linux (x64 + ARM64)
RequirementsDocker
UnlocksOpenClaw view in sidebar

Building Your Own Skills

Skills are Python packages that follow a standard interface. If you know Python and have experience with ML models, you can create your own:

  1. Study an existing skill from the SharpAI GitHub as a template
  2. Follow the standard skill structure (entrypoint, model loading, frame processing)
  3. Host it in a Git repository
  4. Add that repository URL in the Skill Source Browser (gear icon in Skills panel)
  5. Your skill appears in the Marketplace and can be installed like any other

The skill specification supports configurable parameters, multiple platform targets, and custom capabilities. Community-created skills go through the same install, start, stop, and remove lifecycle as official skills.