App Store Operator

App Store competitive intelligence, inside Claude.

App Store Operator is an MCP server that brings App Store research directly into your AI

assistant. Instead of switching to a dashboard, you ask Claude for ranked keyword results,

competitor download and revenue estimates, or App Store Connect-ready In-App Event copy —

and get the answer in the same conversation where you are making the decision.

Built for indie iOS developers who want research inside their workflow rather than in

another browser tab. Free and open source (MIT). A lightweight alternative to SensorTower,

AppTweak, and AppFollow for iOS-only competitive research.

claude mcp add --transport stdio app-store-operator -- npx -y app-store-operator@latest

→ app-store-operator.com · Setup guide

What it does

Searches the App Store for competing apps on a given keyword and pulls detailed analytics

from SensorTower — downloads, revenue, ratings, top markets, publisher info, and more.

`search_app_store` and `prepare_iae` work with no account at all. `research_rivals` and

`get_app_details` open a browser once for a free SensorTower sign-in, then reuse that

saved session — no paid plan, no API key.

Everything the server exposes — four tools, six prompts, seven resources — is read-only.

Nothing writes to your App Store Connect account, to SensorTower, or anywhere but a local

cache file.

Tools

`research_rivals`

Finds the top 3 apps for a keyword and returns a full metrics report for each.

Returns for each competitor:

• App Store & SensorTower URLs
• Worldwide and last-month downloads & revenue
• Rating score and rating count
• Publisher, categories, top markets
• Release date, last updated, supported languages
• In-app purchases and ad network presence

Cached for 24 hours, so asking again about the same keyword and country costs nothing and

opens no browser.

`search_app_store`

Searches the App Store for a keyword and returns ranked results as a markdown table — instantly, no SensorTower required.

Use this to discover which apps rank before deciding which to analyse. Follow up with `get_app_details` for analytics on specific apps.

`get_app_details`

Fetches SensorTower analytics for one or more app IDs you already have.

Returns for each app:

• Downloads and revenue (worldwide + last month)
• Rating score and rating count
• Publisher, categories, top markets
• Release date, last updated, supported languages
• In-app purchases and ad network presence

Never cached — every call scrapes fresh, at roughly 10–20 seconds per app ID.

`prepare_iae`

Generates iOS App Store In-App Event (IAE) copy — 3 variations in the target language, then a final report.

Returns: a structured brief used to generate 3 copy variations, each with event name (≤30 chars), short description (≤50 chars), and long description (≤120 chars).

Any field SensorTower does not expose, or keeps behind its paywall, comes back as `N/A`.

The server reports the gap rather than filling it, and the prompts below tell the assistant

to do the same.

Prompts

Six ready-made workflows that already chain the tools above, so you don't have to describe the sequence yourself. In Claude Code they appear as slash commands; other clients surface them in a prompt picker.

Arguments marked `?` are optional. Every prompt tells the assistant not to invent figures and, where SensorTower is involved, not to quietly fall back to a weaker tool when login is required.

Resources

Reference data and local state a client can attach as context without spending a tool call on it.

Nothing here leaves your machine: the reference resources are static, and the two cache resources read `~/.app-store-operator/cache.json`.

Requirements

• Node.js v18+
• A desktop session for `research_rivals` and `get_app_details`. They drive a real,

visible Chromium window so you can sign in to SensorTower, so they need a display —

they do not work over plain SSH or inside a container. The other two tools have no

such requirement.

• Disk space for Chromium. Installing the package downloads a Playwright Chromium

build (a few hundred MB) via a postinstall step. If that step fails, the server installs

it on first use instead; you can also run `npx playwright install chromium` yourself.

Usage

As an MCP server (Claude Code / Claude Desktop / OpenAI Codex)

Claude Code — run this command once:

claude mcp add --transport stdio app-store-operator -- npx -y app-store-operator@latest

Claude Desktop — add to your MCP config:

{
  "mcpServers": {
    "app-store-operator": {
      "command": "npx",
      "args": ["app-store-operator@latest"]
    }
  }
}

OpenAI Codex — run this command once:

codex mcp add app-store-operator -- npx -y app-store-operator@latest

Codex stores MCP servers in `~/.codex/config.toml`. If you prefer to edit it directly:

[mcp_servers.app-store-operator]
command = "npx"
args = ["-y", "app-store-operator@latest"]

# Optional but useful for SensorTower scraping flows
startup_timeout_sec = 20
tool_timeout_sec = 180

Then restart Codex or start a new thread, and ask things like:

• `Research rivals for "hairstyle" in the GB App Store`
• `Search the App Store for "beard style" in France`
• `Prepare an in-app event for a summer hairstyle campaign in en-gb`

No installation step needed — `npx` fetches and runs the package automatically.

The server communicates over stdio and is designed to be invoked by an MCP client. It advertises server-wide `instructions` during `initialize` so clients route between the tools correctly, and returns an MCP tool error when SensorTower login is required.

Configuration

Both settings are optional environment variables on the server process.

In a container

The repo ships a Dockerfile built on Playwright's official image:

docker build -t app-store-operator .
docker run -i --rm app-store-operator

The server speaks JSON-RPC over stdio, so no port is exposed — point your MCP client at

the container's stdin/stdout. Note that a container has no display: `search_app_store`

and `prepare_iae` work there, but the two SensorTower tools cannot open a login window and

will fail rather than prompting you to sign in.

How it works

1. Searches the App Store for the keyword and country, and looks up any app IDs you passed

directly against Apple's public iTunes Lookup API

2. For each app, drives a Chromium browser to scrape SensorTower analytics

3. Extracts the metrics and returns a compiled report

SensorTower data is scraped via Playwright because it is rendered client-side.

A browser window will open. This is deliberate, not a bug: SensorTower requires a login, so the first run opens a visible window for you to sign in. The session is saved to `~/.app-store-operator/profile` and reused on every later call, so you only log in once. If a tool reports `not_logged_in`, finish signing in on that window and run the tool again.

Results from `research_rivals` are cached for 24 hours in `~/.app-store-operator/cache.json` — override the TTL with the `ASO_CACHE_TTL_HOURS` environment variable.

Limitations

• iOS only. Nothing here covers Google Play or Android.
• Read-only. No tool changes anything in App Store Connect or on SensorTower.
• `research_rivals` is fixed at the top 3 results. Use `search_app_store` (up to 25)

and then `get_app_details` when you need a wider set.

• Scraping is brittle by nature. SensorTower renders its dashboard client-side and

changes its markup without notice; when it does, affected fields return `N/A` until the

selectors are updated. A single app failing never fails the whole call.

• SensorTower's free tier decides what you see. Paywalled figures come back as `N/A`.

Privacy Policy

Full policy: ****

App Store Operator runs entirely on your machine. There is no backend, no telemetry, and no analytics — the author collects, receives, and stores nothing about you or your usage.

What each tool sends, and where:

What is stored locally:

• `~/.app-store-operator/cache.json` — cached results, expiring after 24 hours by default (`ASO_CACHE_TTL_HOURS`)
• `~/.app-store-operator/profile` — the Chromium profile holding your SensorTower session

You type your SensorTower credentials into SensorTower's own page in a browser window on your machine. The server never reads or stores your password, and the author never receives it.

Deleting everything — no request to the author, nothing to wait for:

rm -rf ~/.app-store-operator

Apple and SensorTower are independent third parties with their own policies. This project is not affiliated with either.

Project structure

src/
├── index.js                    # MCP server setup and request handlers
├── shared.js                   # App Store lookup + SensorTower scraping
├── cache.js                    # 24h local cache (research_rivals only)
├── prompts.js                  # the six prompt workflows
├── resources.js                # reference data + cache resources
└── tools/
    ├── research-rivals.js      # research_rivals tool
    ├── search-app-store.js     # search_app_store tool
    ├── get-app-details.js      # get_app_details tool
    └── prepare-iae.js          # prepare_iae tool
scripts/postinstall.js          # installs Playwright Chromium on install
scripts/sync-version.js         # syncs server.json, manifest.json and CHANGELOG.md on release
test/smoke-test-mcp.js          # stdio smoke test
server.json                     # MCP registry manifest
manifest.json                   # Claude Desktop / MCPB bundle manifest
Dockerfile                      # container build (no display: search + IAE tools only)

Development

No build step and no linter — clone it, `npm install`, and run `npm start` to boot the

server over stdio.

The smoke test verifies the server boots and exposes everything it should. It checks

`initialize` (including the server instructions and the advertised version), `tools/list`,

`prompts/list`, `prompts/get`, `resources/list`, `resources/templates/list`, and reads

every resource, failing if one declared as JSON does not parse. It makes no network calls

and opens no browser:

npm run smoke

Releases are tag-driven: write the notes under `## Unreleased` in `CHANGELOG.md`, run

`npm version ` — which syncs the version into `server.json` and

`manifest.json` and renames that heading to the new version for you — then push with

`--follow-tags`. GitHub Actions runs the smoke test, publishes to npm and the MCP

registry, and creates the release from that changelog section. The bump refuses to run

while `## Unreleased` is empty.

Contributions are welcome — open an issue or a pull request at

github.com/meyusufdemirci/app-store-operator.

License

MIT © Yusuf Demirci