Meet Argent: Give Your AI Agent Eyes on Your Mobile App

If you have ever built a mobile app with an AI coding agent, you know the loop: the agent writes some code, you run the app, you take a screenshot of what shows up on screen, and you paste it back so the agent can see what happened. It works. But it is slow, manual, and it snaps you out of flow every few minutes.

Argent closes that loop. It is an open-source agentic AI toolkit from Software Mansion that gives your AI agent direct control over iOS simulators and Android emulators — so it can launch your app, tap through it, read the logs, and profile performance on its own, without you playing messenger.

Modern language models are genuinely good at writing React Native and Swift. Point one at a feature request and it will produce plausible, often correct code. What it cannot do out of the box is see the result. It has no idea whether the button it just wired up actually renders, whether the navigation lands on the right screen, or whether a list scrolls at 60 FPS or stutters.

So we fill that gap by hand. We run the app, we describe what we see, we copy console errors back into the chat, we screenshot a janky animation and hope the agent can reason about it. Every one of those steps is a context switch, and every context switch is where momentum leaks away. The agent is doing the writing; you are doing the seeing. Argent hands the seeing to the agent too.

What Argent is

Under the hood, Argent is two pieces working together:

A set of agent skills — focused, well-described capabilities the agent can invoke, from launching a build to attaching a debugger.
An MCP server — the bridge that connects the agent to running iOS simulators and Android emulators and streams real feedback back.

Because it speaks the Model Context Protocol and runs as ordinary shell commands, Argent works with any agent that can execute a terminal — Claude Code, Cursor, GitHub Copilot, Google Gemini, and others. There is no bespoke plugin per tool; if your agent can run a command, it can drive your app.

What it can actually do

Argent breaks down into three things a real developer does every day: interact, debug, and profile.

1. Drive the simulator

Argent launches an app, then taps, swipes, types into fields, and navigates using coordinates — with real-time feedback after every action. The agent does not guess whether a tap worked; it sees the resulting screen and decides what to do next. That is the difference between a script and an agent: the loop is closed.

2. Debug like a developer

When something breaks, the agent can attach the debugger, explore the view hierarchy, and read console logs. It can inspect React components and monitor network requests — the same signals you would reach for in Flipper or the React DevTools, now available to the agent directly instead of relayed through you.

3. Profile and optimize

This is where Argent goes further than a simple UI robot. It can record both React and native iOS profiles simultaneously and trace a slow React commit down to the native stack frame. Instead of "the screen feels laggy," the agent gets an actual cause — which is what it needs to propose a real fix rather than a guess.

Getting started

Argent targets the Apple toolchain, so you will need macOS, Xcode, and Node.js 18+. From the root of your project, run:

npx @swmansion/argent init

That sets up the skills and the MCP server and wires them into your agent. From there, you can ask your agent to build and launch the app, walk a user flow, reproduce a bug, or profile a screen — and it will do it against a live simulator, reporting back what it sees at each step. The first time you watch an agent open your app, tap through onboarding, hit an error, read the stack trace, and fix it without you touching the simulator, the workflow clicks.

Who it is for

Argent is built for teams shipping React Native and Swift apps who are already leaning on AI agents and are tired of being the manual feedback channel. If you are pairing with Claude Code or Cursor on a mobile codebase, Argent removes the most tedious part of that partnership — the endless screenshot-and-paste — and lets the agent verify its own work.

The bigger picture

Agentic coding tools have mostly given models the ability to write. The next step is giving them the ability to check — to run what they built and observe reality. Argent is Software Mansion's take on that step for mobile: give the agent eyes on the simulator, close the loop, and let it iterate the way a developer actually does.

Explore the docs and see it in action at argent.swmansion.com.