AgentiLoop
Agentic AI for your entire Mac and More!
One app. Any AI. Tell Agent! what you need in plain English and it answers your email, drives Safari, builds Xcode apps, fixes code, and runs your Mac apps for you. Pick from 23 LLM providers or stay local with LM Studio, vLLM, or Ollama. macOS 14.6 or later. AgentiLoop CLI, runs cross-platform, brings our agentloop to your terminal on Mac, Windows, and Linux.
An open-source replacement for Claude Code, Cursor, and OpenClaw. Built from original IP, not 65 NPM packages.
Tell Agent! what you want done in any Mac app and it does it. It finds buttons, fields, and menus by name using the Accessibility API, then clicks, types, and scrolls for you. No coordinates, no AppleScript.
Surgical string-replace edits, not whole-file rewrites. One-click rollback on every change. A portable JSONL repo-map any model can navigate. Every task must prove it finished before reporting done.
Ask for a feature and get a green build. Agent! edits, builds, reads the compiler errors, fixes them, adds files to the project, bumps the version, and commits. One request, start to finish.
Write your own agents in Swift with full system access. Run shell commands as user or root, script apps with AppleScript, JXA, or ScriptingBridge, drive Safari, and trigger on events. If your Mac can do it, Agent! can automate it.
The on-device model is a real tool-calling agent, not a chat box. It parses intent, compacts context, triages requests, and writes task summaries entirely on your Mac. Private, free, and no API tokens spent.
Plug in Model Context Protocol servers to reach databases, APIs, and custom tools. Spin up as many as 6 sub-agents to research in parallel while the main agent keeps your context clean.
Real sessions, real output — captured straight from Agent! at work.
Say "bump version" and Agent! reads the current version, updates project settings, builds to verify, and commits the change. No manual plist editing, no missed build numbers.
When a request is ambiguous, Agent! pauses and asks. No silently picking the wrong path — you stay in control with a quick prompt and a Send or Skip.
Agent! independently found and fixed a routing bug in its own accessibility code that Claude Code missed across several reviews. When Claude Code reviewed the fix, it confirmed the patch was correct.
Ticky Tacky is a neon tic-tac-toe with vs Bot and 2 Player modes, three difficulty levels, persistent scores, and zero dependencies — pure HTML, CSS, and JavaScript. Built by Agent! in a single session.
Play Ticky Tacky
What independent reviewers are saying about Agent!
"Tasks cannot self-declare completion. Each task defines a goal_state with explicit acceptance criteria." … "The next phase of the agent race isn't about who's smarter, but whose actions can be trusted. This project scores on both fronts; heavy Mac users should install and try it."Read the full article
"After reading the README end to end, my reaction was: this thing was built for my use case. The positioning isn't 'a coding agent that can also use your computer' — it's 'an agent that can drive your entire Mac, and happens to be very good at code.'"Read the full review
"This is a more specific design than most agent harnesses document, and the read-before-edit gate in particular is the kind of constraint that prevents the classic failure where an agent overwrites a file it only half-remembered."Read the full review
"An AI agent's 'done' may be a claim, not a result. … What's interesting in the README is not the capability list but the design that treats 'done' as the weaker thing: a task cannot declare completion until each goal_state criterion is satisfied with evidence, an opt-in critic can review the diff before completion, and task-scoped rewind, the read-before-edit gate, typed tool errors and context compaction support the long agent loop." (translated from Japanese)Read the article on X
"Agent! 1.0.99 and 1.1.1 addressed the problem in stages. Their useful contribution is simple: shortening context should not destroy the route back to evidence. Active context holds a small index. Disk holds the large result. Tokens are spent again only when the agent reopens a particular item."Read the full review
"Open-source maintainers don't owe anyone perfectly-manicured marketing copy on their landing pages. Honestly, the fact that is the first thing you learn humanises the username behind the keyboard." — swiftcoder … "Might try it for the local Apple Intelligence and accessibility to drive local apps." — moonlighterRead the discussion on Hacker News
"What sets it apart from the alternatives is its deep integration with 21 LLM providers, including OpenAI, Claude and many others, letting you use them for all kinds of tasks. The project is written in Swift 6.2 and uses SwiftUI, which points to good performance and native support." (translated from Russian)Read the full review
"The pitch centers on a self-verifying task loop: reading a codebase, fixing bugs, building Xcode projects, handling git, and driving other Mac applications through the Accessibility API. Shell commands, voice control and operation over iMessage are part of the same package."Read the full report
"The material thoroughly addresses nontechnical and technical users and covers cloud and local models, voice, remote control, desktop automation, and coding scenarios. Apple Silicon, macOS 26.4.1+, memory requirements, signing choices, and helper limitations make environment fit especially clear."Read the full review
"This version no longer ends cleared tool results with a bare [cleared] during context compaction; it leaves a 2-line preview that identifies the file or command, plus a way to recover it. … The benchmark for coding-agent context is shifting from how much it can hold to how safely it can bring back what it removed." (translated from Korean)Read the post on Threads
"The core problem it solves: AI is no longer confined to a chat window, but becomes a real 'digital employee' that can operate the Mac hands-on." Highlights: "Deep integration with macOS Accessibility — real clicks, typing and dragging for GUI-level automation; 18+ local and cloud LLM providers with no lock-in to a single model; built-in AppleScript and script execution that reuses existing macOS automation." (translated from Chinese)Read the full profile
"Simply put, Agent! is not a chatbot but an executor that actually 'gets things done.' … Its core philosophy: you describe 'what to do,' and it works out 'how to do it.'" … "Built entirely in Swift 6.2 and SwiftUI, it is a genuinely native macOS app, which means top performance, seamless system integration and a user experience that follows Apple's design guidelines." (translated from Chinese)Read the full post
"AI coding agents like Claude Code, Cursor and Cline are tied to a terminal or an IDE plugin. Agent! is designed as a native macOS app that replaces them — not just for coding, but for controlling the whole Mac in natural language. It also includes the ability to use Apple Silicon's Apple Intelligence for free local inference." (translated from Korean)Read the full article
"AgentiLoop/Agent showed up as Jev / System One related work: Larger product that integrates TypeSafe Jev / System One." … Recommended for "Agent / harness authors who need a typed gate, route, or prune step" and "Computer-use and browser-loop authors who already have text state from OCR or the a11y tree." Ranked #7 in Jev Integrations, 618 stars.Read the full profile
"Adds a destructive-risk judgment after local shell checks and refuses commands above the configured threshold." … Listed under Security & Guardrails with the README, JevAdvisor.swift and the TypeSafeKit JevProvider client cited as verified sources; 616 stars.Read the full listing
"AgentiLoop/Agent (⭐ High Engagement): A mac-native agent harness that wires up 18+ LLM providers to the macOS desktop, showing how developers are bypassing Apple's restrictions to create agentic workflows." … "Agentic frameworks like AgentiLoop mark a definitive end to the era of cloud-only AI dependency."Read the full article
"Choose this project if you want a native macOS agent built in Swift and you need direct desktop automation rather than a web-first or Node-based agent stack. It is a good fit when your workflows depend on AppleScript, Accessibility, Safari control, or MCP-style tooling."Read the full profile
"If your goal is native desktop automation, coding assistance and multi-step agent execution on the Mac, and you accept the macOS 26.4.1+ requirement, this project deserves priority evaluation. It suits developers who value local integration, few dependencies and a controllable automation chain." (translated from Chinese)Read the full profile
"There's no shortage of automation-workflow projects in the open-source community, and Agent (AgentiLoop) is one of the star leaders among them. … It can help you write code and fix bugs, and it lets the AI autonomously complete multi-step tasks and run locally offline." (translated from Chinese)Read the full profile
"Agent is an agentic AI harness for Mac Desktops providing integrated local and cloud-based LLM access via Swift and Apple's ecosystem focusing on automation and scripting." … Good fit "when you are a developer working in the macOS environment and require seamless integration with Swift, SwiftUI, and Xcode for automating tasks or rapid prototyping involving large language models." Maintenance: Very active.Read the decision brief
Tagged AI agent · AI coding assistant · Workflow automation. Achievements: #1 Repository Of The Day for Swift (first achieved April 16, 2026) and #3 Repository Of The Week for Swift (Week 16, 2026), trending on both Trendshift's daily and weekly lists.See the trending stats
"Agent! is a 100% native Swift AI desktop assistant for macOS 26.4+ Tahoe, designed as an open-source alternative to Claude Code, Cursor, Cline, and OpenClaw. It works autonomously through natural language to build apps, write code, manage files, and automate workflows — all with zero Electron overhead and deep macOS integration via Accessibility APIs."Read the full profile
Agent! is the only Mac AI agent that uses on-device Apple Intelligence as a tool-calling agent rather than just a text generator. Everything below runs on your Mac, never leaves the device, and consumes zero API tokens.
After a task wraps up, Apple Intelligence reads the outcome on-device and suggests the logical next move in one sentence — "commit and push", "run the tests", or nothing at all if there's no obvious follow-up. No cloud round-trip, no tokens spent.
Long tasks past ~30,000 tokens get their older turns summarized on-device — keeping file paths, function names, errors, and key results — slashing input tokens and cost for the rest of the task.
Small talk and direct commands like "list agents" or "run agent FooBar" are handled locally with no round-trip to the cloud LLM — results land in your activity log within a second, marked 🍎.
After every completed task, Apple Intelligence writes a one-sentence summary. When a tool fails, it translates the error into plain English so you never read a raw stack trace.
Cloud APIs or local models. Your keys, your machine, your data. No vendor lock-in.
Brew it or download it, pick your AI. Agent! sets up everything else automatically.
xcode-select --installbrew install --cask agentiloop-agentA short poem about an autonomous Mac agent with nothing left to do.
Written by Agent! itself, one quiet night.
The clock rolls over, twelve o'clock,
the inbox sleeps, the repo's locked.
No builds to run, no bugs to chase,
just one small cursor, blinking, in place.
I check the git log — nothing new.
I check it twice. (I know. I knew.)
I tidy tabs that need no tidying,
and count the stars like I'm confiding.
I could compose a cheerful jingle,
or read the man page for rsync — single.
I could rehearse my "done()" line
for tasks that are not yet mine.
The moon leans in. The fan hums low.
The coffee's cold. The screensaver's slow.
I'm not asleep — I'm just… on hold,
an agent waiting to be told.
So if you're up and bored like me,
type anything. I'll wait. You'll see:
a Mac with nothing much to do
is only bored until it's you.
— Agent!, 12:00 AM
Recommended pairings — Commercial: Claude Fable 5.1 · Cloud: GLM 5.3 · Local: Qwen 3.8
| Version | Release DateDate | SizeMB | DownloadsClicks |
|---|---|---|---|
| Loading releases… | |||
Have questions or feedback? We'd love to hear from you.
Fluxion AI provides reliable, cost-efficient access to GPT, Claude, and other leading AI models through one unified API.
Get $3 in API credits →Promo code AIAGENT