🦾 AgentiLoop Agent!

Agentic AI for your Mac Desktop

README Translations

• English
• Español
• Français
• Deutsch
• 中文 (简体)

Chess within Agent

What is Agent!?

One app. Any AI. Total command over your Mac.

Agent! is a 100% native Swift 6.2 / SwiftUI app that wires 18 LLM providers — Claude, GPT, Gemini, Grok, Mistral, DeepSeek, Qwen, Z.ai, BigModel, Hugging Face, OpenRouter, Ollama (cloud and local), vLLM, LM Studio, Codestral, Mistral Vibe, and on-device Apple Intelligence — into an autonomous task loop that actually does things: reads your codebase, fixes the bug, builds the Xcode project, commits the diff, drives any Mac app through the Accessibility API, runs shell commands as you or as root, texts you results over iMessage, and answers to a spoken "Agent!".

No NPM, no Electron, no subscription, no telemetry. Bring your own API key, run fully local, or run free on Apple Intelligence. Every Swift package it depends on was written by the same author. See Backstory below.

What's New 🚀

v1.1.x — The Hardened Harness Release · Releases →

• Context compaction, rebuilt. Threshold = model window − reserved output − buffer, driven by real input_tokens. Provider-side 9-section LLM summary replaces on-device 4K summaries; open goal, plan checklist and edited files are re-attached after every compaction. Oversized tool results are spilled to disk at emission and recoverable via restore_tool_result. 413 overflow routes through forced compaction with a shorter retry; max_tokens overruns recover by escalating, then continuing.
• Read-before-edit gate. edit_file / apply_diff / diff_apply refuse to touch a file the LLM hasn't read this task, or that changed on disk since the last read (SHA-256). The refusal auto-reads the file so the next call is the edit. External file changes are surfaced each turn as diff snippets.
• Real context windows for local models. LM Studio, Ollama and vLLM report their actual per-model context length — no more hardcoded 32K assumption.
• Faster turns. Read-only tools start while the Claude response is still streaming; input-aware shell concurrency; jittered exponential retry with Retry-After on 429/529; mid-stream SSE errors surfaced on every provider.
• Defense-in-depth. ShellSafetyService is now enforced daemon-side (AgentHelper + AgentUser) as well as client-side; release builds reject un-teamed XPC clients; both XPC listeners require same-team code signing derived from the app's own signature.
• Activity log. No more 50K truncation or 500K relaunch trim — large logs render off-main with a "Processing tab data…" overlay; optional "Activity Log Below HUD" layout.
• App menu: Check for Updates… (GitHub releases), Website, GitHub. CI Build & Test workflow on every PR; 273 passing tests.
• Plus: goal_state with evidence-verified criteria, opt-in critic diff review before completion, task-scoped rewind_task, extended thinking for Claude, reasoning_effort pass-through, sub-agents with per-agent model override (3 concurrent, 6 read-only), typed tool errors with recovery hints, event hooks.

Quick Start (Download)

1. Download Agent! and drag to Applications
2. Open Agent! — it sets up everything automatically
3. Pick your AI — Settings → choose a provider → enter API key

Quick Start (Build from Source)

git clone https://github.com/AgentiLoop/agent.git
cd Agent

Option A — Xcode (Apple Developer account): open Agent.xcodeproj, set your Development Team, Build & Run the Agent target, approve the helper when prompted.

Option B — no developer account (Xcode Command Line Tools only):

./build.sh              # Debug
./build.sh Release      # Release
open "build/DerivedData/Build/Products/Debug/Agent!.app"

⚠️ Option B builds are ad-hoc signed. The Launch Agent/Daemon helpers won't register (SMAppService needs a Team ID), but the LLM loop, all tools, Accessibility, AppleScript, shell, and MCP still work.

💡 Cheap setup: GLM-5.1 via Z.ai (fastest signup, default model) costs pennies per million tokens. Running locally? Only GLM-4.7-Turbo (32B) fits consumer hardware (64–128GB Apple Silicon via Ollama).

Troubleshooting (Build from Source)

• xcode-select points at Command Line Tools → sudo xcode-select -s /Applications/Xcode.app/Contents/Developer
• Odd BUILD FAILED after pulling → stale DerivedData: ./build.sh clean && ./build.sh
• Helpers never register → expected on Option B; use Option A for the helpers
• Deployment target / SDK errors → Agent! targets macOS 26; update macOS and Xcode
• Config argument is case-sensitive → ./build.sh (Debug) or ./build.sh Release

What Can It Do?

"Build the Xcode project and fix any errors" · "Play my Workout playlist in Music" · "Take a photo with Photo Booth" · "Send an iMessage to Mom saying I'll be home at 6" · "Open Safari and search for flights to Tokyo" · "Refactor this class into smaller files" · "What calendar events do I have today?"

Just type what you want. Agent! figures out how and makes it happen.

Key Features

• 🧠 Self-verifying task loop — reasons, executes, observes results, self-corrects. A task can't declare itself done until goal_state criteria are marked with evidence; an opt-in critic reviews the diff first.
• 🛠 Agentic coding — reads codebases, edits with string-replace diffs, builds Xcode projects natively (clickable errors), manages git, indexes repos into a portable JSONL repo-map. Every edit is snapshotted — one-click rollback or whole-task rewind_task.
• 🖥 Desktop automation — drives any Mac app through the Accessibility API (AXorcist), element-based with fuzzy auto-retry. Plus NSAppleScript, JXA and 51 ScriptingBridge app bridges, all in-process with TCC.
• 📜 AgentScript — Swift dylibs compiled at runtime and dlopen'd in-process with full TCC. Deleted scripts go to .Trash and are restorable.
• 🛡 Privileged execution — shell as you via a Launch Agent, or as root via a Launch Daemon you approve exactly once (SMAppService + XPC). See docs/SECURITY.md for why SMAppService already enforces signing identity.
• 🎙 Voice — say "Agent!" followed by your task; on-device SFSpeechRecognizer, auto-runs after ~2.5s of silence, loops.
• 📱 iMessage remote control — text Agent! next song from your iPhone; approved senders only. Needs Full Disk Access for chat.db.
• 🌐 Web — built-in Safari automation (JavaScript + AppleScript); optional Selenium and Playwright MCP for cross-browser.
• 🤝 Sub-agents — up to 3 concurrent (6 read-only) isolated agents with mailbox messaging and per-agent model override.
• 🧩 MCP — add any MCP server in Settings → MCP Servers; tools appear as mcp_<server>_<tool>. Xcode MCP: {"mcpServers":{"xcode":{"command":"xcrun","args":["mcpbridge"],"transport":"stdio"}}}.
• 🗂 Tabs, history, memory, plans, skills — each tab has its own project folder and log; persistent user memory; multi-plan checklists surfaced in every prompt.
• 🔄 Fallback chain — auto-switch to the next configured provider on 429/timeout/network failure.

🤖 18 AI Providers

💡 Self-hosted providers are free only in the API-fee sense — a usable 30B+ model needs an M2/M3/M4 Ultra Mac Studio (64–128GB). Without that hardware, the cheap cloud paths above are dramatically cheaper.

Tools

Canonical names come from