Most conversations I have about MCP start with the same misunderstanding: people expect the model itself to do something. It can't, not on its own. The Model Context Protocol is the layer that closes that gap — it lets an AI assistant reach something outside the chat window, whether that's an Android phone or an app store account. Below is how the pieces fit together and how I'd wire up a first server without getting lost.
A brain and a pair of hands
The Pony project puts the idea better than I can: Pony is the hands, you pick the brain. Your model decides what should happen; the MCP server is what actually executes it on the device or account. Once I started thinking in those two halves, most of the setup confusion disappeared. The model you pick sets how well the agent reasons. The server you pick sets what it is allowed to touch.
That split also explains why "which model supports MCP?" is the wrong first question. Claude, Grok, Gemini and any other MCP-capable agent can drive the same server — you are choosing a brain, not a platform lock-in.
The checklist before you touch anything
- An AI brain: Claude, Grok, Gemini, or any MCP agent
- Hands: a server that performs real actions
- The target ready — an Android phone, App Store Connect, Google Play Console
- An MCP client that can launch and connect to servers
- Credentials for whatever external service you're connecting
- A clear-eyed acceptance of the trade-off: you are handing an agent direct access
That last point is not a formality. Once permissions are granted, the server does what the model asks. Decide up front what you're comfortable with.
Three servers worth knowing
- Pony — hands an Android phone to an AI agent as a remote control. It needs a physical device.
- shipstores — bridges App Store Connect, Google Play Console and Expo EAS, so an agent can build and publish iOS and Android apps.
- magaza-mcp — a single server covering both stores, with the entire interface in Turkish.
Pick the one that matches your problem before you start installing anything.
Wiring it up, step by step
- Choose the brain. Start with a model that speaks MCP. This decides the quality of the decisions, not the reach.
- Choose the hands. Pony for a device, shipstores or magaza-mcp for store publishing.
- Install the server on the target side. For Pony that means running it on the Android phone itself; for the store servers, a machine that can reach your store accounts. You're done when the server is up and waiting for a connection.
- Register it in your client. Open the configuration area of your AI client and add the server address or startup command. You're done when the client discovers the server and lists its available tools.
- Grant permissions. Pony needs accessibility or comparable device-level access. The store servers need your App Store Connect and Google Play Console accounts connected, plus the API keys those accounts require.
- Run a small test. Keep it boring: "open the camera" or "check the latest app build status." If a real answer comes back from the device or the store, the connection works.
What usually goes wrong
The agent can't see the server. In my experience this is almost always ordering. Start the server first, then open the MCP client — a server that comes up afterwards often goes undetected. Restart both and double-check the path in your config.
Nothing happens, silently. On Android that usually means missing device-level permissions. With the store servers it means the accounts aren't properly connected or a key is absent. A silent failure is nearly always a credential failure.
The server answers, but with store errors. For shipstores, confirm your account's access level actually allows publishing. For magaza-mcp, remember the interface is Turkish throughout — a language mismatch can mask an account-local problem.
Quick answers
Is one server enough for both stores? Yes — shipstores covers App Store Connect and Google Play Console together, and magaza-mcp does the same in a single server.
Do I need a real phone for Pony? Yes. Giving an agent an Android smartphone over MCP is the entire point of the project.
Does the client matter? It matters as a launcher: without a client that can start and connect to MCP servers, the server just sits there.
If you want to dig further into agents that use this protocol in day-to-day work, there is a longer write-up on AI Release covering automation agents — https://ai-release.org — and new guides land on the channel at https://t.me/ai_release1.
Takeaway
MCP is not magic and it is not complicated: it is a contract between a brain that decides and a server that acts. Choose both deliberately, install the server before you open the client, grant only the access you actually need, and validate with one trivial instruction. If "open the camera" returns a real result, you have connected your first MCP server — and everything after that is just adding more hands.
Top comments (0)