MCP server · CLI · Watch daemon

Let any coding agent see your screen

ScreenSight is a privacy-first screen capture layer for coding agents. It ships as an MCP server for Cursor, Windsurf, Cline, Claude Code and Codex CLI, a plain CLI for everything else, and a bounded watch daemon for capture over time. No API key. Off by default.

Get started View on GitHub

pip install .
screensight on
screensight capture

Start here

Installation Install from source, with pip, or with pipx. Python 3.10+ on Windows, macOS, Linux and WSL. Install →

Quickstart Enable the switch, take your first capture, and confirm your agent can see the frame — about two minutes. Two-minute tour →

Agent setup Copy-paste MCP config for Claude Code, Cursor, Windsurf, Cline, Codex CLI — plus the CLI path for Aider. Wire up your agent →

MCP tools All eight tools the server exposes, with inputs, outputs and example responses. Read the reference →

CLI reference Every command, flag, JSON payload and exit code for the screensight binary. Browse commands →

Privacy model Master switch, sensitive-app blocklist, zone redaction, frame hygiene and watch self-limits. See the guarantees →

Why ScreenSight

Agent-agnostic by construction

One MCP server works in every MCP-capable agent with zero agent-specific code. Agents without MCP shell out to the CLI instead.

Off unless you say otherwise

A master switch gates every capture, and the check lives inside the capture pipeline — not in a tool description an injected instruction could argue with.

Sensitive windows never land on disk

Password managers, keychains and incognito windows are matched by title and abort the capture before a frame is written.

One frame, not a shoebox of screenshots

A single frame.jpg is overwritten on every capture and deleted when you turn the switch off.

Bounded watching

The watch daemon stops itself after a frame budget and flips the master switch off on exit, so an idle daemon can't quietly keep burning tokens.

Cheap frames

Every image is downscaled to a 1568px long edge — past the point extra pixels help a vision model — keeping token cost low for every consumer.

How it compares

Feature	ScreenSight	ScreenPipe	Agent-specific tools
Works with any MCP agent	✅	❌	❌
CLI fallback (no MCP needed)	✅	❌	Varies
Watch daemon with auto-stop	✅	✅	❌
Master switch (off by default)	✅	❌	Varies
Blocklist + zone redaction	✅	❌	❌
No API key required	✅	✅	Varies
Cross-platform	✅	✅	Varies
1568px downscale for tokens	✅	❌	❌
Frame cleanup on off	✅	❌	❌
Open source	✅	✅	Varies

Inspired by ScreenPipe — this is the agent-agnostic version.