AgentHandover

I've been using Claude Code, Cursor, OpenClaw -- all of it. And the thing that keeps driving me crazy is that none of them remember how I work. I'll set up a perfect workflow, nail the prompting, get great output... and it forgets. Back to square one. Explaining the same context, the same preferences, the same guardrails. Also sometimes I just wish it would know what to do without me even saying anything.

So I spent the last few months building something to fix that.

AgentHandover sits in your Mac menu bar and watches you work. Not in a creepy way -- it runs entirely on your machine, local models only, zero cloud, zero telemetry. It captures what you're doing, figures out the patterns, and packages them into Skills that any AI agent can pick up and run.

There are two modes. You can hit record and deliberately teach it something, or you can just let it run in the background and it'll find your recurring workflows on its own. The second one is honestly kind of wild. It'll surface patterns you didn't even know you had.

Under the hood there's an 11-stage pipeline doing screen capture, VLM annotation with Qwen 3.5, activity classification, semantic clustering, cross-session linking, behavioral synthesis, voice analysis... all of it crunching locally through Ollama. Screenshots get deleted right after processing.

The output is a Skill file. Not a vague summary. An actual structured playbook with your strategy, your decision logic, your guardrails, your writing voice. And it has a confidence score that goes up when agents execute successfully and goes down when they don't. The Skill literally improves itself.

Every Monday morning I pull numbers from four different places. GitHub stats, website analytics, social engagement, and email subscriber counts. I drop them into a running spreadsheet, compare to last week, note anything unusual, and write a short summary for myself so I can spot trends over time. It's not hard work. It's just boring and I keep putting it off until Tuesday or Wednesday, which defeats the purpose.

AgentHandover watched me do this four Mondays in a row. It learned which sources I pull from, which numbers I care about, how I format the spreadsheet entries, and what counts as "unusual" in my head (basically anything that's up or down more than 15% from the previous week). The agent now runs the roundup Monday at 9am and drops a summary on my desktop. The confidence started low because my "unusual" threshold wasn't consistent -- sometimes I'd flag a 10% change, sometimes I'd ignore a 20% one. After I corrected a few outputs it settled at 83% and the flags are now more consistent than my own judgment was.

Stack is Rust + SwiftUI + Python + TypeScript Chrome extension. Pairs with Claude Code, OpenClaw, Codex, Cursor, Windsurf, anything that speaks MCP. Apache 2.0 license. Hit 104 stars in under 24 hours.

Install takes a minute. Download the .pkg, grant permissions, pull models, done.
CLI native for devs and agents. 

Once click connect to any agent be it OpenClaw, Codex or Claude Code, etc.

Star it: https://github.com/sandroandric/AgentHandover Download: https://www.agenthandover.com

What workflow would you hand over first?