phone-eye
Enables AI agents to see and operate a real Android phone over adb by fusing live screenshots with UI-tree data, supporting look, tap, swipe, type, and screenshot actions through any MCP client.
README
Phone-Eye š±šļø
Let your AI agent see ā and operate ā a real Android phone.
Your coding agent can read code, run commands, and browse docs. Can it see the app it's building, on a real phone, with real pixels? Phone-Eye gives it eyes and a finger.
phone_look("what's on screen, and where is the login button?")
ā "Login button at (540, 1830) ā a green 'Sign in' ā¦"
phone_tap(540, 1830)
phone_look("did the next page load?")
Five verbs, no framework: your agent composes them into whatever workflow it needs. Works with any MCP client ā Claude Code, Codex, Cursor, dsh, and friends.
Why
If you build or test anything that ends up on a phone, you know this loop: the layout is broken on the real device, you screenshot by hand, describe screens in words ("the gear, top right, next to the account thing"), and play coordinate-decoder between your agent and your phone. Phone-Eye closes that loop ā the agent iterates with the device the way it already iterates with your codebase.
Two channels, fused:
- Vision ā a vision model answers natural-language questions about the live screenshot (works on game canvases, images, anything pixels can show).
- UI tree ā
uiautomatordump for exact text and bounds when the accessibility tree has them.
Vision is pluggable. Any MCP server exposing describe_image(path, question) works ā cloud GLM vision out of the box, or point it at a local
Qwen-VL endpoint so no pixel ever leaves your LAN.
Install
Requirements: Python 3.10+, adb (platform-tools / android-tools) on PATH,
a phone with USB debugging on, and any MCP vision server.
git clone https://github.com/boheastill/phone-eye
cd phone-eye
pip install -r requirements.txt
# env (defaults shown):
export ANDROID_SERIAL="" # empty = first adb device; or 192.168.x.x:5555
export PHONE_EYE_VISION_URL="http://127.0.0.1:8102/mcp" # your vision MCP
Wire it into your client (stdio):
// Claude Code / dsh-mcp-client / any stdio MCP config
{
"mcpServers": {
"phone-eye": { "command": "python", "args": ["/path/to/phone-eye/server.py"] }
}
}
Or run it as an HTTP service (streamable-http) behind your own fleet and add
http://<host>:8122/mcp ā see docs/fleet.md.
Connect the phone (one-time)
USB once, then Wi-Fi forever:
adb devices # USB: accept the debugging prompt on the phone
adb tcpip 5555 # switch to Wi-Fi mode
adb connect <phone-ip>:5555 # unplug and go
Tools
| Tool | What it does |
|---|---|
phone_look(question?, use_tree?) |
Ask a vision model about the live screen; fuses UI-tree text + bounds |
phone_tap(x, y) |
Tap |
phone_swipe(x1, y1, x2, y2, ms?) |
Swipe |
phone_type(text) |
Type ASCII (spaces ok; CJK needs clipboard route ā known adb quirk) |
phone_screenshot() |
Save screenshot to disk, return path |
What it is / isn't
ā agent eyes + hands on one real Android device, zero on-device install, no root ā vision-first (works where UI trees can't see) with tree-fusion for precision ā privacy option: point vision at a LAN-only model
ā not a test framework (no DSL/recorder ā the agent is the logic) ā not iOS, not device farms (yet) ā not a mobile UI for humans (that's a different product)
Status & roadmap
Phase 1 (now): the five verbs, single device, pluggable vision ā verified end-to-end on real hardware (Redmi K40 Gaming / Android 13). Its favorite party trick so far: it discovered a USB-debugging authorization dialog on its own screen, read the buttons, and tapped "Allow" itself.
- [ ] Phase 2: offline vision quick-start, multi-device addressing, retry/verify wrappers
- [ ] Everything else: request-driven ā open an issue and it moves up the queue.
License
MIT
FAQ
Vision server? I don't have one. Any MCP server exposing
describe_image(path, question) works. The quickest cloud option is a GLM
vision endpoint; for fully-offline, a local Qwen-VL (llama.cpp / Ollama
OpenAI-compatible + a 20-line adapter) keeps every pixel on your LAN.
Why not a dsh-native plugin? MCP-first means the same five verbs work in
every client. dsh users can wire it via @deepseek-ai/dsh-mcp-client or
dsh plugin add.
Verified on
| Device | Android | Connection | Notes |
|---|---|---|---|
| Redmi K40 Gaming (ares) | 13 (HyperOS) | Wi-Fi adb (adb tcpip 5555) |
daily driver of the author's fleet |
Add yours via a PR to this table.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.