ghostwork: Your Personal AI Worker That Watches, Learns, and Automates

The 'AI that watches you work and learns to do it' idea has been promised for years - Adept AI raised $400M on this premise, Rabbit R1 launched and disappointed. Most of these projects died.

ghostwork takes a different approach: it's open source, runs locally, and learns from screen capture + user confirmation. You do a task 3-5 times, ghostwork watches, then proposes a workflow. You approve, and it runs the workflow next time.

How it works in practice

Install ghostwork, give it screen recording permission, and do your normal work. The first time you do a task, it asks 'should I learn this?'. You confirm, do the task with ghostwork recording. Next time, ghostwork proposes the same steps and asks 'should I run this?'. The user-in-the-loop design prevents runaway automation.

Use cases that work today

Data entry from PDFs to spreadsheets. Email triage and template responses. Web form filling. Code formatting. These are the bread and butter of personal automation, and ghostwork handles them competently.

What doesn't work yet

Complex multi-app workflows (e.g., 'take this Slack message, create a Linear ticket, and schedule a meeting') are still hit-or-miss. The model is small and local, so it can't reason about novel situations the way GPT-4 can. This is more 'macro recorder with smarts' than 'true AI employee'.

Verdict

A genuinely interesting 2026 prototype. The open-source, local-first design is the right call. If you have 5+ hours/week of repetitive screen work, ghostwork will pay back the setup time. Just don't expect magic - this is RPA with an LLM, not AGI.