AgentiLoop/Agent

What it solves

Agent! is an autonomous AI agent for macOS that allows users to control their entire desktop environment through a single interface. It eliminates the need for manual app switching and repetitive tasks by enabling an AI to read code, execute shell commands, drive Mac applications via the Accessibility API, and manage system-level operations natively.

How it works

The app is built with Swift 6.2 and SwiftUI, creating a native macOS experience. It connects to 18 different LLM providers (including local options like Ollama and on-device Apple Intelligence) and puts them into a self-verifying task loop. The agent reasons through a goal, executes tools, observes the results, and self-corrects until a defined goal state is reached. It uses a combination of Accessibility API for UI automation, AppleScript/JXA for app control, and a custom "AgentScript" system that compiles Swift dylibs at runtime to perform complex system tasks with full TCC permissions.

Who it’s for

Developers and power users on macOS who want to automate their workflow, manage codebases, and control their system using an AI agent that can operate both locally and via cloud APIs.

Highlights

• Broad AI Provider Support: Integrates with 18 providers including Claude, GPT, Gemini, and local models via Ollama or LM Studio.
• Agentic Coding: Capable of reading codebases, editing files with diffs, and building Xcode projects natively.
• Deep macOS Integration: Controls apps via Accessibility API, runs shell commands (as user or root), and integrates with iMessage and Voice commands.
• AgentScript: A system for creating and running Swift scripts that are compiled at runtime for high-performance, native system access.
• Self-Verifying Loop: Tasks are not marked complete until evidence-verified criteria are met, with an optional critic review for diffs.
• MCP Support: Compatible with the Model Context Protocol (MCP) to extend tool capabilities via external servers.