A new open-source project called Agent! landed on GitHub this week. It wants to replace paid AI coding tools with a single native macOS app that actually does things on your computer. Not just suggests code. Actually drives apps, runs commands, edits files.
Built in Swift 6.2, it plugs into 17 different LLM providers including Claude, GPT, Gemini, DeepSeek, and on-device Apple Intelligence. Agent! can drive Mac apps through Accessibility APIs. It can run shell commands as root via a Launch Daemon. Execute Xcode builds. Send iMessages. Respond to voice commands. Every file edit gets a Time Machine-style rollback point. Free. No telemetry. No provider lock-in.
The scope is ambitious for what appears to be a solo project. Version 1.0.69.158 already includes autonomous task loops that reason and self-correct, MCP server support, prompt caching across OpenAI-format providers, and on-device token compression using Apple AI to keep context windows manageable. You can run it fully local through Ollama, vLLM, or LM Studio. Or use Apple Intelligence and pay nothing, similar to how Apfel exposes the model hiding on your Mac.
The README pitches it as a direct alternative to Claude Code, Cursor, Cline, and OpenClaw.
There's a catch. Hacker News commenters flagged the Launch Daemon feature as a potential security concern. It lets the agent execute root-level commands after a one-time approval. Giving an LLM that level of system access requires trust in both the model and the middleware. The project also requires macOS 26.4 or later, which limits the audience.
The founder, who goes by toddbruss on GitHub, is battling cancer and has asked for community support through stars and forks. If you're on a Mac and want a self-hosted AI coding assistant that goes beyond chat, Agent! is the most interesting option right now. Just think hard before handing it root access.