Clousd MCP

What clousd-mcp MCP server does

The clousd-mcp MCP server gives an MCP client access to real Android devices hosted by Clousd. It is intended for agents and automation that need to see and operate mobile app interfaces. Devices represent specific handset models and include the Android system, Google Play services, cameras, sensors, and network characteristics described by the provider.

The server supports interactive work as well as repeatable runs. An agent can inspect the current screen, choose an element or coordinate, perform an action, and check the resulting state. Saved device snapshots can be restored so repeated tasks begin from a known condition.

How it works

The client communicates with Clousd over MCP. A local installation uses stdio; the hosted option uses the streamable endpoint at https://api.clousd.com/mcp. Authentication is supplied with a Clousd API key. The hosted server does not retain state between requests, according to the README, so the device and authorization scope are carried by the request context.

Screen observation returns a screenshot, the foreground app, numbered elements, and a UI tree containing text, identifiers, and bounds. Actions include taps, long presses, swipes, scrolling, text entry, key presses, app and link opening, intents, settings pages, clipboard operations, and the notification shade. Actions can wait for the screen to settle, and stale observations are rejected with a 409 stale response.

Setup and configuration

Install the local package with:

pip install clousd-mcp

Configure an MCP client to launch clousd-mcp and provide CLOUSD_API_KEY in its environment. Alternatively, configure the client with the hosted URL and an Authorization: Bearer ... header. Access keys have read or control scope and can be restricted to an explicit device list. Read-only keys can observe and inspect but cannot perform actions.

Access to the service is described as early access, with keys requested through Clousd's agents page. The README does not document additional local configuration variables.

Tools and capabilities

The clousd-mcp MCP server is suited to tasks that require:

• Observing a phone screen together with its accessible UI structure.
• Acting on Android interfaces through coordinates or numbered elements.
• Opening apps, links, intents, and settings pages.
• Reading notifications, clipboard contents, app versions, crashes, ANRs, and health information without changing the screen.
• Restoring a device to a saved snapshot before a run.
• Using an ADB relay for deeper checks, with a one-time code, IP allow-list, and expiry.
• Running recipes and fleet operations through the MCP fleet capability, where available in the Clousd account.

The broader Clousd platform also provides recipes for recorded or scripted app flows, app packs, device cloning, schedules, jobs, webhooks, and parallel execution across up to 32 phones in one call. These capabilities are part of the platform described by the repository and may depend on account access or use through the API and Python client.

Limitations and notes

Clousd access requires an API key, and the repository identifies the service as early access. The MCP package does not create a local Android emulator; it connects to hosted real-device infrastructure. A read-scoped key cannot tap, type, or otherwise control a phone.

The README documents the hosted URL and local stdio mode, but it does not provide a complete list of MCP tool names or a detailed account-pricing model. ADB access is separately guarded by a one-time code, IP allow-list, and expiration. Use fresh observations before acting because stale screen sequences are refused.

The clousd-mcp MCP server is therefore a fit for Android UI automation, agent evaluation, and repeatable mobile workflows when access to Clousd's devices is available. It is less suitable when the requirement is an entirely local device runtime or an emulator-only test environment.