mrkeyoor.com_
Tue 01 Sept 17:41 UTC
Dev Toolsevaluationupdated 26 Aug 2026

argent review

Argent is an MCP toolkit that lets coding agents control, inspect, record, and profile apps on iOS simulators, Android devices, TVs, Electron, and Chromium. It gives Claude Code and other supported agents device actions such as tapping, taking screenshots, reading UI trees, replaying flows, and collecting diagnostics.

+81stars / 7d
Verdict

Our Argent run installed 501 packages, built in 30 seconds, and exposed 11 known vulnerabilities while providing no test target, so it needs a security and behavior review before an agent touches a real app. Its device breadth is useful for an experienced mobile team that wants interactive diagnosis inside Claude Code or another MCP client. Do not make recorded flow passes a release gate until the false-pass issues are fixed and your own failure cases behave correctly.

We ran it

Lab card: what happened when we ran argentScreenshot of argent (argent.swmansion.com)
Install✓ · 47s501 packages · 401 MB
Build✓ · 30s
Testsn/ano test script
Known vulns110 critical · 8 high · 3 moderate · 0 low (npm audit)
Repo1210 files~233,095 lines of source · 84 MB · 18 CI workflows

Answers from our run

Does argent build from source?

Dependencies installed in 47 seconds (501 packages), and the build succeeded in 30 seconds. We cloned commit 2740a3e into a clean Debian container with 3 CPUs and no project-specific setup.

Does argent have tests you can run?

Not through a standard command: the project exposes no test script or target that our harness could run.

Does argent have known vulnerabilities in its dependencies?

npm audit flagged 11 known advisories in the dependency tree at the time of our run.

Who should not use argent?

Organizations requiring every distributed component under an open-source license: the README says several simulator, accessibility, and native-devtools binaries are proprietary and restricted.

What are the alternatives to argent?

Maestro, Appium, Playwright. Our Argent run installed 501 packages, built in 30 seconds, and exposed 11 known vulnerabilities while providing no test target, so it needs a security and behavior review before an agent touches a real app.

Setup3/547-second install; device hosts need substantial platform setup
Docs5/5Targets, host prerequisites, install modes, and telemetry are clear
Community4/52,246 stars with heavy August 2026 pull request activity
Maturity2/5Build passes, but no test target and false-pass reports remain

Who it’s for

Mobile teams that want Claude Code, Codex, Cursor, or another MCP client to reproduce bugs on simulators and Android devices.
React Native developers who need agent-accessible UI trees, logs, network inspection, and profiling.
Test engineers willing to review recorded flows and visual results before accepting an agent's pass.
Teams that can standardize Xcode, Android SDK, emulator, and platform-specific host setup across developer machines.

Who it’s NOT for

Organizations requiring every distributed component under an open-source license: the README says several simulator, accessibility, and native-devtools binaries are proprietary and restricted.
CI owners who need a visible project test target before adoption: our checkout had no test script or tests directory.
Teams that treat a green recorded flow as proof: issues 964 and 965 show failure-shaped tool results being recorded or scored as passes.
Visual-regression pipelines that cannot audit baseline identity: issue 961 describes scoped crops sharing a baseline and potentially passing against the wrong image.
Developers unwilling to grant an agent control over device input, cookies, storage, JavaScript evaluation, network capture, and local profiling tools.
Privacy policies that forbid opt-out diagnostics: telemetry is enabled unless it is disabled through the documented command or local installation option.

Setup reality

Our sandbox installed 501 npm packages in 47 seconds and used 401 MB. The build succeeded in 30 seconds. The repository had no test script or target, so we skipped tests. Npm audit reported 11 known vulnerabilities: 8 high and 3 moderate.

The base install needs Node.js 20.12 or later. iOS and tvOS require macOS and Xcode; Android needs adb and emulator packages; Fire TV needs the Vega CLI. Electron and Chromium targets must expose a Chrome DevTools Protocol port.

Linux Android emulators also need working KVM access for usable speed, a suitable system image, and optional headless settings. Local installs build native tree-sitter dependencies on each machine, with a C or C++ toolchain required on targets lacking prebuilt packages.

One MCP server can control 4 very different interface families

Argent gives an AI coding client a common tool surface for iOS simulators, Android emulators and physical devices, TVs, and Electron or Chromium applications. The actions change with the target: touch gestures on phones, remote buttons on TVs, and mouse or keyboard input on desktop apps. An agent can also inspect component trees, logs, network activity, cookies, storage, screenshots, and profiler output. That breadth is the reason to consider Argent. It is also a large permission grant.

The installer registers Argent as an MCP server for Claude Code, Cursor, VS Code, Windsurf, Zed, Gemini CLI, Codex CLI, Hermes, OpenCode, or Kiro. It copies agent instructions and project files into the workspace. A local mode can pin Argent in devDependencies and commit the generated configuration so teammates receive the same version. Global mode is quicker for one developer, though it makes project behavior depend on whatever version is installed on that machine.

A 30-second build passed without any project test target

Our Node 22 sandbox installed 501 npm packages in 47 seconds and used 401 MB. The build completed in 30 seconds. The checkout contained 1,210 files, about 233,095 source lines, and 84 MB before dependencies. It is a workspace monorepo with 18 CI workflow files and no Dockerfile. Those figures make the clean build encouraging, especially for a package that connects to several native toolchains.

Tests were a different story: the repository exposed no test script or target, so our harness skipped them. The scan also found no tests directory. We did not substitute an internal command or infer coverage from the 18 workflow files. A successful TypeScript build confirms compilation at commit 2740a3e; it does not prove that taps, replays, tree reads, screenshot comparisons, or profiler sessions produce correct verdicts on supported devices.

What happened when we ran it

Our install and build took 77 seconds combined on 3 CPUs with 8 GB of RAM. Npm audit reported 11 known vulnerabilities in the installed dependency tree: 8 high, 3 moderate, and 0 critical or low. The supplied measurement does not identify the packages or show whether a vulnerable path is reachable. A team should inspect the full audit result before letting the MCP server handle app credentials, cookies, storage, captured traffic, or arbitrary JavaScript evaluation.

We did not boot an iOS simulator, connect Android over adb, launch a TV image, or attach Chrome DevTools Protocol to an app. Those workflows require host hardware and system packages outside the fresh Debian container. Our run therefore measures repository setup and compilation only. Argent's most important promises concern real devices and stateful app interaction, so a useful evaluation must replay known passes and known failures on the same target types your team ships.

Recorded flows can report work as passing when it failed

Issue 964 shows flow-add-step recording a debugger status as a successful step even when the result says no app is connected. The next replay then scores the same step as failed. Issue 965 expands the class: an idle wait can lose its last error, a device boot can return a native-devtools initialization failure, and a Chromium cookie set can return false while the flow still appears green. These reports are code-specific and were open on August 25, 2026.

Issue 961 describes a visual-regression collision. Two screenshot crops with the same name and basic selector, but different relational scopes, can receive the same baseline identity. Updating baselines may overwrite the first image with the second, and a later run can compare against the wrong crop without warning. For test automation, these are serious defects because they affect the verdict rather than merely the display. Keep human-readable screenshots and negative controls in the acceptance suite.

Platform setup is part of the product, especially on Linux

Argent requires Node.js 20.12 or later. iOS and Apple TV work needs macOS, Xcode, and the relevant simulator runtimes. Android requires Platform Tools, adb, and an emulator package if Argent should boot virtual devices. Fire TV work adds the Vega SDK. Electron and Chromium control uses Chrome DevTools Protocol, either by attaching to a process launched with a debugging port or by letting Argent start it.

Linux emulators need access to /dev/kvm; otherwise they fall back to software emulation. The project chooses SwiftShader by default to avoid broken or black windows on mixed graphics stacks, with an environment override for verified host rendering. Headless runners need a no-window setting. Project-local installation compiles native tree-sitter dependencies when prebuilt binaries are unavailable, so Linux arm64 and Windows arm machines also need a C or C++ compiler.

Apache-2.0 source ships with proprietary native binaries

The repository metadata names Apache-2.0, and the README applies that license to the source code. It separately says per-platform simulator-server binaries, the macOS accessibility service, and native iOS devtools libraries are proprietary Software Mansion components. Their terms prohibit decompiling, reverse engineering, or redistribution without permission. That mixed model may be acceptable for internal use, but it is materially different from an all-Apache toolkit for vendors that redistribute developer environments.

Telemetry is opt-out, with commands to inspect or disable it. Release v0.22.1 added a project-scoped opt-out and a local installation flag for disabling telemetry. GitHub showed 2,246 stars, 286 combined issues and pull requests, and a last push on August 26, 2026. The active queue suggests rapid correction as well as rapid change. Argent is worth a guarded mobile-team trial; the 11 advisories, absent test target, mixed license, and false-pass reports rule out blind release-gate use.

Alternatives

ProjectWhat it isPick it when
MaestroA declarative mobile UI testing framework for Android and iOS flows.pick this instead when repeatable mobile test files matter more than giving an agent a broad debugging toolbox.
AppiumA long-running automation server built around WebDriver-compatible mobile testing.pick this instead when your team already uses WebDriver clients and needs conventional test-runner integration.
Playwright gh↗A browser automation and testing framework with tracing and cross-browser support.pick this instead when Chromium, Firefox, and WebKit testing is the job and native mobile control is unnecessary.

What people are saying

  1. [github-trending] software-mansion/argent

Sources

  1. Argent repository and README
  2. Argent v0.22.1 release
  3. Failure results scored as passing
  4. Debugger status recording issue
  5. Scoped visual baseline collision
  6. Argent telemetry notice

More dev tools reviews

workmux · v2rayNG · SecLists · hashcat · eslint · fastfetch · the whole board →