activity-frames

Why we included this project

Computer-use agents tend to re-derive the same repetitive tasks from scratch, and between sessions they have no memory of what you've been doing all day. activity-frames tackles both problems by recording your screen locally and compressing the raw captures into compact activity frames: deterministic records of what you actually did, including the app, site, timestamps, and input counts. Those frames feed two outputs. One is a short context block you can drop into a system prompt so your agent knows your day. The other is a set of executable steps for recurring workflows, which the agent can replay instead of reasoning through the task again. Because everything is measured from screen capture rather than guessed, the compiled output costs almost no tokens to run and stays consistent. If you build agents that operate on a real desktop, or you're just tired of agents starting every session blind, the CLI and MCP server give you a concrete way to turn observed work into repeatable actions.