screen-mcp
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@screen-mcpcapture the screen and annotate clickable elements"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
screen-mcp
Linux Wayland computer-use MCP for Claude Code and Grok — screenshot, click, type, scroll, drag, read any desktop app.
screen-mcp is an MCP server and plugin for Claude Code and Grok for desktop
automation on GNOME/Wayland. It captures any monitor through PipeWire, drives
pointer and keyboard via the xdg-desktop-portal RemoteDesktop portal, and
optionally reads the screen with OCR plus OmniParser icon grounding. Pure
Python, CPU-only — built for AI agents that need real computer-use on native
Linux apps, not just a browser.
It ships the MCP tools plus a drive-screen skill that encodes the locate →
ground → act → confirm/watch loop so the model uses screenshots and clicks
well out of the box. screen_watch (default ~1 fps) is how agents see like
humans: thrashing graphs, loaders, and canvases fail visual QA with a jitter
verdict instead of slipping past a single freeze-frame. Developers get
productivity tools that reach every visible window — including native Wayland
apps that xdotool and XTEST cannot touch.
Install
Claude Code
/plugin marketplace add 88plug/claude-code-plugins
/plugin install screen-mcp@88plugGrok Build
grok plugin marketplace add 88plug/claude-code-plugins
grok plugin install screen-mcp@88plug --trustThen install system packages (once per machine) and Python deps (plugin venv). The marketplace cannot install these for you.
# --- system (pick your distro) ---
# Debian / Ubuntu
sudo apt install python3-gi python3-gi-cairo gir1.2-gstreamer-1.0 \
gstreamer1.0-tools gstreamer1.0-plugins-base gstreamer1.0-plugins-good \
gstreamer1.0-libav pipewire pipewire-pulse xdg-desktop-portal-gnome \
wl-clipboard fonts-dejavu
# Arch / Manjaro
sudo pacman -S python-gobject gobject-introspection \
gstreamer gst-plugins-base gst-plugins-good gst-libav \
pipewire pipewire-pulse xdg-desktop-portal-gnome \
wl-clipboard ttf-dejavu
# --- Python (required: numpy + Pillow; optional OCR later) ---
# From the installed plugin root (or a clone):
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
# minimum only: .venv/bin/pip install 'numpy>=1.26' 'Pillow>=10.0'Related MCP server: desk-mcp
Quickstart
On first use the desktop portal asks which monitor(s) to share. Pick one, then:
Take a screenshot of my desktop and tell me which window is focused.You get a labeled capture plus the focused-window name within a few seconds. The
portal restore token is cached at ~/.config/mcp-screen so later runs are silent.
screen-mcp runs on Linux + Wayland + GNOME only. Grounding is CPU-only by design. SeeRequirements before installing.
Features
Feature | Detail |
Screenshot + Set-of-Marks | Capture any monitor or region with numbered overlays and click coordinates |
Click / type / scroll / drag | Drive any visible app over xdg-desktop-portal, including native Wayland |
OCR + icon grounding | Optional RapidOCR text read and OmniParser ONNX icon grounding |
Ambient change sense | Frame diffs so the agent knows when something opened or an action no-op'd |
World-model cache | Write-through screen memory skips OCR on recognized UIs |
Ack guard | Opt-in gate blocks close-combos and destructive-keyword clicks until confirmed |
| Claude skill for the locate → ground → act → confirm computer-use loop |
Principles — The Agent Oath
screen-mcp is a reference enforcer of The Agent Oath.
The user-takeover guard yields the instant a human moves the mouse (STOPPED
result) — §2 human agency and §11 human oversight, executable. Opt-in ack gate
(§7) and on-screen visibility of every action (§5) round it out.
MCP tools
Every action also accepts space (view / desktop / norm, default view —
coords as seen in the last screenshot), shot: true for a post-action shot,
verify: true to warn on no-change misclicks, force: true to bypass the
user-takeover guard, and element: <id> to click an element id from the last
annotated shot.
Tool | What it does |
| Capture the desktop. |
| Monitors (origin/size/scale), desktop bounds, focused windows. |
| Move the pointer to |
| Click at |
| Wheel scroll by direction and amount. |
| Press-drag from one point to another. |
| Press a key or combo, e.g. |
| Type text into the focused window (Unicode via clipboard paste, ASCII via keysyms). |
| Read the screen as text + click coords, no image (3.5x-50x fewer tokens). |
| Block until text appears, then return its click coords. |
| Grade the last action: CONFIRMED / PARTIAL / NO_OP / DIVERGED. |
| Copy the focused window's selection and return it verbatim — exact text, no OCR. |
| Raise and give keyboard focus to a window by app/title/id. |
| Run a batch of ordered actions in one call. |
| Visit several UI states, return a labeled thumbnail of each. |
| Auto-scroll a scrollable view and accumulate every interactable. |
| Block until the screen settles, then optionally screenshot. |
| Recorder: |
| Hot-reload the server in place after edits. |
| Health dump: session, cursor, grounding backends, world-model stats. |
| Normalized cross-layer change signal from the last frame diff ( |
Requirements
screen-mcp targets a specific stack. It will not run elsewhere.
Linux + Wayland + GNOME. The awareness layer uses a bundled GNOME Shell extension; AT-SPI is the fallback for GTK apps.
Python 3.10+ runtime (tested on 3.14). CI pins 3.12 because optional the OCR backend is
rapidocr(v3), which supports 3.8-3.x.GStreamer >= 1.28 (uses
leaky-type; the olderdrop=was removed in 1.28).PipeWire and
xdg-desktop-portal-gnome.wl-clipboardfor the Unicode paste path inscreen_type.A DejaVu Sans Bold font for Set-of-Marks labels (falls back to PIL's default).
Grounding is CPU-only by design: the server hard-disables the GPU
(CUDA_VISIBLE_DEVICES="") for predictable latency and no driver flake.
Install the system deps before the Python deps. See
requirements.txt for the full pacman / apt one-liners.
# Arch / Manjaro
sudo pacman -S python-gobject gobject-introspection \
gstreamer gst-plugins-base gst-plugins-good gst-libav \
pipewire pipewire-pulse xdg-desktop-portal-gnome \
wl-clipboard ttf-dejavu
# Python deps
pip install -r requirements.txtThe bundled window-info extension gives the awareness layer reliable
focused-window and window-list data, and lets screen_focus activate windows
directly. Installing it needs a one-time Wayland re-login.
gnome-shell-extension/window-info@local/install.sh
gnome-extensions enable window-info@localFor the kernel input backend, grant access to /dev/uinput by adding your user
to the input group. The launcher (bin/screen-mcp) fails with a clear message
if a required dep is missing, so a misconfigured install never silently
half-works.
Manual MCP setup
If you are not using the plugin, wire the server via the launcher (resolves host
Python / .venv through scripts/run-python.sh — never bare python3 on thin
Claude spawn PATH). Add to ~/.claude.json under mcpServers:
{
"screen": {
"command": "/path/to/screen-mcp/bin/screen-mcp"
}
}Configuration
screen-mcp reads optional environment variables. A few common ones:
MCP_SCREEN_GUARD=1— enable the ack gate. Destructive combos (Ctrl+W,Alt+F4), OCR-matched destructive keywords, and out-of-allowlist actions block unless the caller passesack=<reason>.MCP_SCREEN_APPS="firefox,terminal"— with the guard on, restrict actions to this allowlist of focused apps.MCP_SCREEN_AMBIENT=0— disable the ambientSENSEhint block.MCP_SCREEN_CPU_THREADS=6— ONNX intra-op thread count for OmniParser.MCP_SCREEN_MAX_EDGE=2576— screenshot downscale target (long edge).
Variable | Effect |
| Enable the reliability ack gate (destructive combos / keywords / out-of-allowlist actions block until |
| With guard on, allowlist of focused apps. |
| Add pre/post frame hash + |
| Disable the ambient |
| Threshold for the user-takeover guard (live pointer vs last-commanded). |
| ONNX intra-op thread count for OmniParser. |
| Screenshot downscale target (long edge). |
| Disable forced fresh-frame capture on static monitors. |
| Delay after |
| Disable the pointer damage-nudge that primes a static monitor's frame. |
Path | What |
| Portal restore token (one-time consent). |
| World-model cache (per-screen learned elements). |
| Recorder trajectories, frames, |
| Reliability audit log (one JSON line per action). |
| Last unhandled tool traceback (dev diagnostic). |
Development
pytest -q # tests run without a live D-Bus (conftest stubs)Edit a .py, then call screen_reload in the running session to re-exec the
server in place while preserving the MCP connection. On any tool exception the
dispatcher writes the full traceback to /tmp/screen_err.txt; read it when
debugging crashes.
Contributing
Issues and pull requests are welcome. Please run pytest -q before opening a PR.
See CLAUDE.md for architecture notes and the hard-won ops details
behind the capture and input layers.
License
FSL-1.1-ALv2 © 2026 88plug. The Functional Source License converts to Apache 2.0 two years after each release.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- AlicenseAqualityDmaintenanceAn MCP server for Hyprland desktop automation that allows AI assistants to see the screen, control mouse and keyboard, and manage windows using native Wayland tools. It integrates OCR for text-based interaction and supports complex multi-monitor setups with pixel-accurate coordinate mapping.Last updated275MIT
- Alicense-qualityDmaintenanceA desktop automation MCP server that enables AI agents to interact with Linux environments through screenshots, window inspection, and input simulation. It provides tools for mouse control, keyboard input, and screen capture using xdotool and XDG Desktop Portals.Last updatedMIT
- Alicense-qualityBmaintenanceAn MCP server that gives any AI assistant eyes and hands on your desktop — screenshots, clicking, typing, OCR, window management, accessibility-tree queries, workflow recording.Last updated5Apache 2.0
- Alicense-qualityDmaintenanceAn MCP server that enables LLMs to see and control a computer — screen capture, window management, mouse and keyboard automation — with a structured plan-execute workflow for complex desktop automation.Last updatedGPL 3.0
Related MCP Connectors
A comprehensive Model Context Protocol (MCP) server that enables AI assistants to interact with yo…
Hosted MCP server connecting claude.ai, ChatGPT and other AI apps to your own computer
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/88plug/screen-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server