For those familiar with idb, some of iosef’s agent-oriented features include:
Directory-level sessions, so agents in different worktrees don’t interfere with each other’s simulators.
Automatic scaling of screenshots into the same coordinate space as the accessibility trees and interaction tools.
Selector-based interactions, to save LLMs from needing to re-consult accessibility trees or screenshots each turn.
Skills and MCP support.
=====
iosef’s absence from the pre-training corpus means agents need a little nudge (either via the skill or user prompt) to use it over idb or simctl—but once they’re using it, they have a easier time than they do with other tools. While written in swift, it’s packaged via pypi for easy running and installation via uv (eg “uvx iosef --help” or “uv tool install iosef”).
Is it nearly entirely vibe-coded? Yes—but I’ve found it helpful for my own use, so I share it in case others will, as well. And this post is entirely human-written, so at least there’s that.
No comments yet.
