Argent
You ask your agent to adjust something on a screen. It changes the code. You open the simulator yourself, tap through the flow, something looks wrong, you take a screenshot, paste it back into chat, and now you’re burning another round just to tell it what went off. That cycle repeats more than it should.
Argent fixes that. It’s a tool from Software Mansion that wires up your coding agent directly to the iOS Simulator and Android emulator through MCP. Once it’s set up, you don’t have to be the middleman anymore.
Getting Started
npx @swmansion/argent init
One command from your project root. The setup wizard finds your editor, registers the MCP config, and drops skills and rules into your workspace. Your agent then knows exactly what it can reach for.
A few requirements before you run it:
- Node 20.11 or newer
- Xcode for iOS Simulator access
-
adbplus an AVD for Android
It works with Cursor, Claude Code, VS Code, and Codex CLI. And it works fully local if that matters to you. A blog post I came across showed it running fine with Ollama and gemma4 on device. No cloud dependency required.
What Your Agent Can Now Do
Run and Navigate
The agent can build your app, launch it, open deep links, tap through screens, type text, and hit hardware buttons. It reads the accessibility tree to know what’s actually on screen, not just what the component file says. Multi step flows work in a single call, with feedback after each action so it doesn’t walk in blind.
Debug
Console logs, crash reports, the live React component tree, and native network inspection. That last one is the part that actually surprised me. There’s an entire layer of native network calls that never show up in your JavaScript logs. If a request got rejected at the native client, Argent surfaces it. The component tree view is what makes this click for React Native specifically since you can see which component owned a bad layout while the bug is still live.
Profile
The agent records React Native profiling sessions and Instruments runs, then returns summaries it can reason about. What’s slow, where the UI thread is taking hits, what’s leaking. You can keep asking questions about the same session without having to record again.