An annotation layer any agent can draw on.
Arin is a local daemon that draws on your screen. Point at things, highlight regions, write explanations. Any AI can drive it over MCP or a small JSON protocol.
It draws. It never clicks.
brew install anistark/tools/arinOr with Nix:
nix run github:anistark/arin -- -dQuickstart · Install · Docs · Website
An agent explaining an article can circle each section as it talks about it. An agent reviewing your code can point at the line it means. A tutorial agent can show you where a button is instead of describing it.
Arin holds no model and makes no decisions. The intelligence is your agent. Arin is the chalk.
Arin never synthesises input. No clicks, no keystrokes, no scrolling. That boundary is the point: a tool that can only render pixels is safe to leave running, needs no Accessibility permission, and is auditable in an afternoon. Agents that need to act should compose Arin with a separate actuator.
No accounts, no analytics, and no network connections with the default settings. Screen
captures stay in memory on your Mac. The two settings that send anything anywhere, the
claude resolver and the update check, are off until you turn them on.
PRIVACY.md has what Arin reads, stores and sends.
MIT