Open source · MIT

Agent Skills

Real devices. Real control. Built for AI.

AI can't test what it can't control.

Give your AI the access it needs to actually test — real Android and iOS devices, real browsers — and the judgement to use it well. Fewer false results, less wasted time.

Installs into Claude Code, Cursor, GitHub Copilot, Codex, Cline, Zed, Amp, Continue and 30+ other agents.

View on GitHub Set up the MCP server

Tools are not instructions

An MCP server tells an agent what it can do. It does not tell it what to do. Handed a device and no guidance, agents tap coordinates read off a downscaled screenshot, type before the field has focus, and report success from a screen they never checked. Every skill here is built around the failure modes that actually happen.

Fewer wasted runs

The agent locates real elements instead of guessing pixels, so a run does not end three screens away from where it should be.

Results you can trust

Assertions read the device. "Passed" means something was verified, not that no tool call threw an error.

Devices freed on time

Sessions get released on the failure path too, so a crashed run does not hold a device against your parallel limit.

Install

First, connect the MCP server

Skills are documentation, not a server. They assume your agent is already connected to RobotActions. If it is not, this one command sets it up — you will need a tenant subdomain and an API token from Portal → Profile → API Tokens.

bash

npx --yes @robotactions/mcp init

Any agent — one command

Served from this domain. Installs every skill in a universal format that Claude Code, Cursor, GitHub Copilot, Codex, Cline, Zed and 50+ other agents read. Nothing else to configure.

bash

npx skills add https://robotactions.com

You do not have to take all of them. Run it without flags and it asks which skills and which agents you want; or preview and pick from the command line:

bash

# see what is available, install nothing
npx skills add https://robotactions.com --list

# install just the ones you want
npx skills add https://robotactions.com --skill mobile-app-testing

Or pull the same skills straight from the source repository:

bash

npx skills add krishtoautomate/robotactions

Claude Code — as a plugin

Prefer a plugin? Add the marketplace, then install it. Same skills, and it registers the MCP server too.

text

/plugin marketplace add krishtoautomate/robotactions
/plugin install robotactions@robotactions

The plugin bundles every skill and an MCP server entry, so installing it also registers the server — against the shared free test tenant, reading your token from ROBOTACTIONS_API_TOKEN. On a paid tenant, point it at your own subdomain with the installer above instead.

By hand

Each skill is a plain folder with a SKILL.md file. Clone the repo and copy the ones you want into wherever your agent loads skills from.

bash

git clone https://github.com/krishtoautomate/robotactions.git
cp -r robotactions/skills/* ~/.claude/skills/

A machine-readable index of every skill, for tooling that wants to enumerate them: skills.json

The skills

Each one scoped to a job an agent is actually asked to do.

Getting started

robotactions-getting-started

Connect an MCP host to the device cloud, authenticate, and prove the connection with a real device before any test work starts.

Teaches

Per-host config shapes — and why Claude Desktop rejects the one Cursor accepts
Reading a 401 as a rotated token instead of patching around it
Device holds, the parallel limit, and releasing on the failure path

Key tools

device_listdevice_infodevice_release

Read the skill

Mobile app testing

mobile-app-testing

Drive real Android and iOS devices through a native app: install it, navigate the UI, assert on what is genuinely on screen, capture evidence.

Teaches

Never inventing a coordinate — and the scaling step that makes taps land
Why a focus tap and a keystroke must never be issued in parallel
Waiting for elements rather than for durations, on both platforms

Key tools

device_tap_by_textdevice_wait_for_elementios_tap_by_labelapp_install

Read the skill

Web app testing

web-app-testing

Test a site in the browser of a real phone or on the desktop grid — responsive checks on real hardware, network debugging, request mocking.

Teaches

Which of the three browser tool families a task belongs to, and what each leaks if unreleased
Snapshot refs instead of brittle CSS selectors
Why a network capture triggered from a separate call records nothing

Key tools

webpage_snapshotwebpage_clickwebpage_mock_addweb_navigate

Read the skill

Device state setup

device-state-setup

Put the device into the state the test needs before driving it — because most "flaky" mobile tests are not flaky, they assume state that is not there.

Teaches

Pushing a file so an upload or file-picker flow has something to pick
Seeding an app's own container with fixtures before it launches
Mocking location and locale — and putting them back so the next run is clean

Key tools

device_file_pushios_file_pushdevice_set_locationdevice_clear_app_data

Read the skill

Network mocking & capture

network-mocking

Intercept and mock traffic for native app requests as well as web pages — error, empty, offline and slow-network states without waiting on a backend.

Teaches

Which of the two interception layers a request belongs to, and why the wrong one silently matches nothing
Decrypting native app traffic that browser tooling cannot see
The two limits that actually break interception, before you blame the rule

Key tools

android_traffic_startandroid_traffic_mock_addios_traffic_mock_addwebpage_mock_add

Read the skill

Framework integration

framework-integration

Use the tools with the suite you already have — Appium, Selenium, Playwright or WebdriverIO — instead of replacing it.

Teaches

Exploring a screen live, then lifting ranked stable locators into page objects
Emitting tests in the framework and conventions the repo already uses
Pointing an existing suite at the grid without rewriting the tests

Key tools

device_locators_forios_find_elementsession_createsession_quit

Read the skill

Flow record & replay

flow-record-replay

Turn one manual walkthrough into a regression test that replays on any device, with assertions that actually fail when behaviour changes.

Teaches

Recording assertions after every state change, not only at the end
Reusable components, so the login flow is written once
Reading a replay that finished green with steps silently skipped

Key tools

flow_recording_startflow_replay_startflow_replay_summarytest_suite_run

Read the skill

Found a failure mode we missed?

These skills are MIT-licensed and written from real runs. If you hit something that is not in them, open an issue or a pull request.

View on GitHub