I built a menu bar app that watches how you work and turns your workflows into self-improving Skills that any of AI agents can execute without you explaining how to do your work. Open source, fully local

Developer Sandro Andric has open-sourced AgentHandover, a novel macOS application that addresses a fundamental limitation of current AI agents: their inability to understand individual user workflows. The menu bar app operates in two modes: Focus Record for one-time task capture and Passive Discovery, which runs for days to classify work activities using an 8-class taxonomy and cluster similar actions. All processing happens locally using a sophisticated 11-stage pipeline.

The technical stack is impressive for a local application. It uses perceptual hashing (dHash) for screen capture efficiency, the Qwen 3.5 2B vision-language model (2.7GB via Ollama) for frame annotation, and nomic-embed-text (274MB) for generating 768-dimensional text embeddings. After observing a workflow three or more times, the system synthesizes behavioral patterns using Qwen 3.5 4B (3.4GB) to create structured Skill files. These Skills include decision logic, step sequences, guardrails, and even capture the user's unique writing voice and tone.

AgentHandover prioritizes privacy and security with a local SQLite vector store, automatic redaction of PII and passwords, encrypted storage (XChaCha20-Poly1305), and zero telemetry. The generated Skills are compatible with any MCP (Model Context Protocol) agent, including Claude Code, OpenClaw, Cursor, and Windsurf. While currently macOS-only with Windows on the roadmap, the Apache 2.0 licensed project represents a significant step toward truly personalized AI assistance that learns from user behavior rather than requiring explicit instruction.