All of Agent!'s IP is original and open source. Every Swift package dependency and the app itself were originally authored by the same person. This is a genuinely different ecosystem. Most agentic AI apps like Claude Code rely on 65 NPM packages from third parties. Agent! is 100% native, requires very little RAM, and weighs in at 35.5 uncompressed. That footprint includes Xcode automation, a Swift Syntax 6.2 package for troubleshooting native apps, Accessibility, AppleScript, AgentScript/ScriptingBridge, Safari automation, MCP Server support, and more. Out of the box.

What's New 🚀

• 🌐 OpenRouter provider: Full integration with OpenRouter's model catalog — one API key, 200+ models. Smart catalog fetches available models, protocol toggle (OpenAI/Anthropic) routes Claude models through the Anthropic protocol and everything else through OpenAI, and ClaudeService auto-routes when OpenRouter is selected.
• Apple AI as a real tool-calling agent: On-device Apple Intelligence (FoundationModels.Tool) handles UI automation requests like "take a photo using Photo Booth" locally — multi-step tool calls, zero cloud LLM tokens, falls through to the cloud LLM only on failure.
• SDEF + runtime app discovery: Bundle ID resolution is now zero-hardcoded. Apps in Agent/SDEFs/ plus every .app in /Applications, /System/Applications, ~/Applications are discovered at runtime — installing a new app extends what the agent can target with no code edit.
• Prompt caching for every OpenAI-format provider: Z.ai, OpenAI, Grok, Mistral, DeepSeek, Qwen, Gemini, BigModel, Hugging Face — cached_tokens is parsed from the response and shown in the LLM Usage panel. JSON request bodies use .sortedKeys so byte-stable prefixes actually hit the provider's cache.
• On-device token compression: Apple AI summarizes old conversation turns when context exceeds 30K tokens (Tier 1 of tieredCompact) — free, private, no API tokens consumed. Toggleable in the brain icon popover.
• Anti-hallucination prompt rule: Every system prompt now includes explicit guidance against fabricating findings from incomplete tool reads. The 10-consecutive-reads guard pushes the model toward "narrow or call done()" instead of "guess".
• Autonomous task loop, Xcode integration, AXorcist desktop automation, privileged daemon, multi-tab LLM config, Ollama pre-warming via LLMRegistry — all the previously-shipped fundamentals are still there.

Star History ⭐️

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

Agent! 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 a native macOS app that doesn't just talk about doing things. It does them.

Watch it read your codebase, fix the bug, build the Xcode project, and commit the diff while you make coffee. Tell it to open Safari and text you the price of flights to Tokyo. Say "Agent!" from across the room and have it run your test suite by voice. Text your Mac from iMessage and get a polished answer before you reach your car.

It edits files with surgical string-replace diffs — every change one-click undoable from a Time-Machine-style rollback. It drives any Mac app through the Accessibility API — no AppleScript required. It remembers your preferences across sessions. It spawns parallel sub-agents for work that fans out. It indexes entire codebases into a portable JSONL repo-map that any LLM can consume. It runs shell commands as you, or as root via a Launch Daemon you approve exactly once.

Bring your own API key. Run it fully local on Ollama, vLLM, or LM Studio. Or run it free, forever, on Apple Intelligence. No subscription. No telemetry. No vendor lock-in. Your keys, your machine, your data.

Download it. Say what you need. Watch it happen.

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)

1. Clone the repository:

   git clone https://github.com/macos26/agent.git
   cd Agent
   

Option A: Build with Xcode (Apple Developer account)

2. Open Agent.xcodeproj in Xcode.
3. Build and Run the Agent target.
4. Approve the Helper Tool: When prompted, authorize the privileged daemon to allow root-level command execution.

Option B: Build without an Apple Developer account

2. Run the build script (requires only Xcode Command Line Tools):

   ./build.sh              # Debug build
   ./build.sh Release      # Release build
   

3. The app lands in build/DerivedData/Build/Products/Debug/Agent!.app
4. Run it: open "build/DerivedData/Build/Products/Debug/Agent!.app"

⚠️ Without a developer account the app is 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 all work.

Then:

5. Configure your AI Provider: Go to Settings and enter your API key or select a local provider like Ollama.

💡 Cheap GLM setup: GLM-5.1 runs on all four cheap providers — Ollama, Hugging Face, Z.ai, BigModel — at pennies per million tokens. New here? Start with Z.ai (fastest signup, GLM-5.1 is the default, nothing to provision). Running locally? Only GLM-4.7-Turbo (32B) fits on consumer hardware (M2/M3/M4 Mac, 64-128GB, via Ollama) — GLM-5 and GLM-5.1 are too large (~1.6TB), use them via the cloud providers above.

What Can It Do?

"Play my Workout playlist in Music" "Build the Xcode project and fix any errors" "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

🧠 Agentic AI Framework

Built-in autonomous task loop that reasons, executes, and self-corrects. Agent! doesn't just run code; it observes the results, debugs errors, and iterates until the task is complete.

🛠 Agentic Coding

Full coding environment built in. Reads codebases, edits files with precision, runs shell commands, builds Xcode projects, manages git, and auto-enables coding mode to focus the AI on development tools. Replaces Claude Code, Cursor, and Cline -- no terminal, no IDE plugins, no monthly fee. Features Time Machine-style backups for every file change, letting you revert any edit instantly.

🔍 Dynamic Tool Discovery

Automatically detects and uses available tools (Xcode, Playwright, Shell, etc.) based on your prompt. No manual configuration required for core tools.

🛡 Privileged Execution

Securely runs root-level commands via a dedicated macOS Launch Daemon. The user approves the daemon once, then the agent can execute commands autonomously via XPC.

Why there's no manual setCodeSigningRequirement on the XPC listener

Users sometimes ask why AgentHelper's XPC listener accepts connections without a manual connection.setCodeSigningRequirement(...) check. The short answer: SMAppService already enforces signing identity one layer below your code, so the check would be redundant.

That recommendation is a holdover from the pre-SMAppService SMJobBless era, where launchd did not validate identity for you and the XPC server had to set a designated-requirement string itself. SMAppService changed that contract:

• The app-bundle-embedded plist plus signature-gated registration is the code-signing requirement.
• The Mach service names (Agent.app.toddbruss.helper, Agent.app.toddbruss.user) are namespaced to the signed bundle that registered them — no other bundle can claim them.
• Any signature mismatch (tampering, re-signing, different Team ID, bundle swap) breaks the XPC channel at the launchd layer — listener(_:shouldAcceptNewConnection:) is never even invoked.

**Empirical pro