AppLab Discovery - MCP Server for Mobile Apps
AppLab Discovery emerged from a practical need: automating the cycle of discovery, testing, and marketing material generation for mobile apps, using AI agents as the primary interface.
The problem that motivated the project
End-to-end mobile app testing involves many manual steps: installing the build, navigating through flows, capturing screenshots, documenting bugs, generating store previews. When I multiplied this across several parallel projects, it became clear I needed an automation layer that an AI agent could orchestrate.
MCP architecture with 15+ tools
The server implements the Model Context Protocol, exposing over 15 tools via JSON-RPC. Each tool is atomic and composable: an AI agent can call a screenshot tool, pass the result to a visual analysis tool, and chain it with a Jira bug creation tool, all in a single conversation.
Tool design follows the single responsibility principle. One tool takes screenshots, another extracts text via OCR, another analyzes accessibility. This allows the agent to compose complex workflows from simple pieces, without the server needing to anticipate every use case.
Multi-platform automation
For web, automation uses Playwright with multi-browser support. For iOS and Android, I integrated Maestro CLI, which allows describing test flows in YAML and executing them on simulators or real devices. The server abstracts platform differences: the agent requests "execute the login flow" and the server picks the correct runner.
Evidence capture is automatic. Each flow step generates screenshots, and video recording can be activated for critical flows. Crashes and ANRs are detected and reported with stack traces when available.
Computer vision and UI analysis
The visual analysis pipeline combines OCR for text extraction with the Claude Vision API for semantic screen interpretation. The AI can identify UI patterns, detect accessibility issues like insufficient contrast or small touch targets, and suggest improvements based on Apple and Google design guidelines.
On Apple devices, the server also integrates the Apple Vision Framework via a native bridge, achieving more accurate OCR results for text in iOS interfaces.
Marketing asset generation
One of the most useful daily tools is automated asset generation for the App Store and Play Store. The server uses FFmpeg to compose promotional videos from screenshot sequences, and Three.js to render 3D device mockups with app screens inserted.
This drastically reduces launch preparation time: instead of manually creating each preview, the agent generates all necessary size and format variations from a base set of screenshots.
Product tool integrations
The server has connectors for Jira, Notion, Figma, and GitHub. Each integration is exposed as an independent MCP tool. The agent can, for example, analyze a screen, identify an alignment bug, create a Jira ticket with screenshot and description, and link the corresponding GitHub issue, all without human intervention.
Local-first security
All execution happens on the local machine. There is no telemetry, no data sent to the cloud, no external storage. API keys are read from environment variables and never persisted to disk. This architecture is ideal for projects with compliance requirements or that handle sensitive user data.