Agent!: a Mac-only agent with 21 LLM providers
Agent! packs 21 model providers into one native macOS app under an MIT license — with every claim so far resting on the project's own repo.
In short
Agent! is an open-source macOS app that routes 21 cloud and local model providers into a single self-verifying task loop, though every detail available so far comes from the project's own repository.
At a glance
- Native macOS app; the repo requires macOS 26.4.1 or later and runs on Apple Silicon only.
- 21 model providers listed, including local runtimes and on-device Apple Intelligence.
- Source code under the MIT license; users pay only their chosen model provider.
- Repo page at time of reporting: 600 stars, 66 forks, 971 commits, 273 passing tests, release v1.1.x.
- The project page is the only available source; no independent testing exists.
Agent! is an open-source agent application that runs only on the Mac and puts 21 model providers behind one interface. The project sums itself up as „One app. Any AI. Total command over your Mac.“ Every claim below traces back to that same repository and nowhere else.
What the project says it does
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. Up to three sub-agents are meant to run concurrently in isolation. External tooling attaches through MCP.
Twenty-one providers, no single lock-in
The roster spans Claude, OpenAI, Gemini, Grok, Mistral and DeepSeek, router services, and local runtimes such as Ollama, vLLM and LM Studio, plus on-device Apple Intelligence. The candidate blurb behind this story still cited more than 18 providers; the repo now says 21. Which figure applied when is not something an outside reader can reconstruct.
Apple-only by design
The app asks for macOS 26.4.1 or later on Apple Silicon and is written in Swift 6.2 with SwiftUI. The project makes a point of shipping without Electron or NPM baggage, with every dependency written in-house. Building from source needs the Xcode command line tools; prebuilt releases are offered as well. The code carries an MIT license, so the only running cost is whichever provider the user picks.
Deep privileges, unaudited guardrails
An agent that holds shell access, Accessibility control and Swift libraries compiled at runtime reaches far into the operating system. Against that, the repo lists shell safety enforcement, a read-before-edit gate, file backups with rewind, and trust checks on XPC clients. Files, screen contents and personal data are said to stay on the machine. None of those assurances has been checked by anyone outside the project.
The open questions
At the time of reporting the page showed 600 stars, 66 forks, 971 commits and 273 passing tests, with the latest release labeled v1.1.x. Those are self-reported figures from a single source. There is no third-party review, no external security audit and no usage data to weigh them against.
FAQ
Which LLM providers does Agent! support?
The repo lists 21, among them Claude, OpenAI, Gemini, Grok, Mistral and DeepSeek, alongside local runtimes such as Ollama, vLLM and LM Studio plus on-device Apple Intelligence.
Does Agent! run on Windows or Intel Macs?
No. The project states macOS 26.4.1 or later on Apple Silicon as a hard requirement, and no other platform is offered.
Is Agent! free to use?
The source is MIT-licensed and the app itself is described as free; the only cost is whatever the connected model provider charges.
