Performance & Latency

What to expect when using ScreenMCP for phone automation.

Overview

ScreenMCP commands travel through a multi-hop relay: your AI client sends a command to the ScreenMCP API, which routes it through a WebSocket worker to the phone. The phone executes the action and sends the result back through the same path. Most commands (tap, type, back, home) complete in under 500ms. Screenshots take longer because of image capture, compression, and transfer.

Connection Setup

The first command in a session requires a connection setup. Subsequent commands reuse the existing connection and are much faster.

Step	What happens	Typical time
Discover	API finds the best worker and notifies your phone	0.6–1.0s
WebSocket + Auth	Client connects to worker and authenticates	0.8–1.2s
Phone wake-up	Phone receives push notification and connects to worker	0.3–2.0s
Total (first command)		2–4s

If the phone is already connected (e.g. you sent a command recently), the wake-up step is skipped. The phone stays connected for 5 minutes after the last command.

Screenshot Performance

Screenshots are the most latency-sensitive command. The phone captures the screen, compresses it to WebP, base64-encodes it, and sends it back through the relay.

Step	Details	Typical time
Screen capture	Android AccessibilityService captures pixels	60–150ms
Compress + encode	WebP compression and base64 encoding	150–300ms
Network transfer	Phone → worker → client (~80KB)	200–500ms
Total per screenshot		0.5–1.5s

Image size

60–90 KB

WebP @ quality 80, 720px wide

Phone processing

~280 ms

Capture + compress + encode

Avg round trip

~900 ms

End-to-end with network

Other Commands

Non-screenshot commands are faster because they don't involve image compression or large data transfer.

Command	Typical round trip
click, long_click	200–400ms
type, get_text	200–400ms
back, home, recents	150–300ms
drag, scroll	300–600ms
ui_tree	300–800ms
screenshot	500ms–1.5s

Tips for Best Performance

1Use lower screenshot quality. Set quality: 60 and max_width: 720 for faster compression and smaller transfer. The AI can still read the screen clearly.
2Keep sessions active. The phone disconnects after 5 minutes of idle time. Sending commands regularly keeps the connection alive and avoids the 1–2 second reconnection delay.
3Use ui_tree for navigation. Instead of taking a screenshot to find a button, use ui_tree to get element positions. It returns text data which is faster to transfer than images.
4Good network matters. The phone needs a reliable internet connection (Wi-Fi or strong cellular). Latency depends on the phone's network speed and distance to the relay server.

How It Works

AI Client (Claude, Cursor, SDK)
    │
    │  HTTPS ── discover worker, get wsUrl
    ▼
ScreenMCP API  ──push notification──▶  Phone
    │                                    │
    │                                    │ connects to worker
    ▼                                    ▼
Worker (WebSocket relay) ◀────WSS────▶ Phone
    │                                    │
    │  command ──────────────────────▶   │ execute
    │  ◀──────────────── response ──    │
    ▼
AI Client receives result

All communication is encrypted (TLS). The worker is a stateless relay — it does not store screenshots or commands. Data flows through in real time and is not persisted.

← MCP Integration TypeScript SDK →