# MCP server runs on Android > Source: > Published: 2026-08-19 15:15:19+00:00 # MCP server runs on Android [MCP](/en/tags/mcp/), an MCP server that runs directly on your phone and exposes tap, type, scroll, and screenshot actions to any compatible agent. I've been testing it for a few weeks now, and the practical implications for AI workflow automation are genuinely interesting. **How it works in practice** The app registers as an accessibility service, which gives it read access to the UI hierarchy across all installed apps. When an agent sends a tool call — say, "open Gmail and search for 'invoice'" — the server parses the accessibility tree, finds the relevant elements, and executes the sequence. Token usage stays low because the server returns structured summaries instead of raw XML dumps. A typical interaction consumes roughly 200-400 tokens per step depending on screen complexity. **Privacy mode is the standout feature** After early feedback made it clear nobody wants their full screen context shipped to an LLM provider, the author added on-device redaction. A small local model combined with deterministic regex detectors strips emails, phone numbers, credit cards, IBANs, national IDs, and English names before anything leaves the device. Benchmarked detection sits around 87% — non-English names remain the weak spot, but a custom model is in development. For anyone building LLM agent pipelines that touch personal data, this approach deserves attention. **Integration with Claude and ChatGPT** The app acts as its own OAuth server. You approve connections via a code displayed on the phone, then point [Claude](/en/tags/claude/).ai, Claude Desktop, or chatgpt.com at the local endpoint. No ngrok, no Cloudflare tunnel — though a free reverse tunnel with Let's Encrypt certificates and true end-to-end encryption is on the roadmap to replace those workarounds entirely. **Tradeoffs worth knowing** - Google Play distribution is impossible because accessibility services trigger policy restrictions. The APK lives on GitHub; a FOSS build without Play Services is heading to F-Droid soon. - Prompt injection mitigation: every response from the server is prefixed with a strong untrusted-input warning. Not bulletproof, but it raises the bar significantly. - App coverage varies. Mainstream apps (Chrome, Gmail, Maps, Spotify) work reliably. Custom or heavily obfuscated UIs sometimes break the tree traversal. **What's coming next** - Skills database per app so agents can plan faster without trial-and-error exploration - Custom privacy model for better non-English name detection - Guided setup flow and revised UI to lower the barrier for non-technical users If you're experimenting with AI workflow automation on mobile, this is worth a hands-on guide session. The repo includes a practical tutorial for getting started from scratch, and even smaller models like Haiku can drive meaningful multi-step tasks when given enough context. [Why most companies will miss the AI-native shift entirely 4h ago](/en/news/6912/) [Claude just fired its first human employee 20h ago](/en/news/6832/) [Free API keys for top models are now public 21h ago](/en/news/6826/) [Finding clean open source alternatives to AI-tainted software is 1d ago](/en/news/6736/) [Will Gen Z actually survive the AI takeover of entry-level roles? 1d ago](/en/news/6721/) [Codex weekly limits are disappearing way too fast for me 1d ago](/en/news/6708/) [Next Anders Hejlsberg just dropped a talk that should make every →](/en/news/6939/)