SkillAgentSearch skills...

kde-mcp

An MCP server that lets an AI agent observe and control a KDE Plasma 6 / Wayland desktop via AT-SPI, KWin Scripting, and libei — accessibility-tree first, screenshots as fallback.

Install / Use

claude mcp add atassis -- npx -y github:atassis/kde-mcp

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

76/100

Supported Platforms

Claude Code
Claude Desktop

Our assessment of kde-mcp

kde-mcp scores 76/100 on our quality scale, 3743rd of 4,716 Development & Engineering skills we index.

Its MCP Server is 5.3 KB long, split into 7 sections with 2 code examples: a solid amount of guidance for an agent.

It has 10 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
26/30
Structure
16/20
Description
15/15
Adoption
4/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated 13 days ago, so kde-mcp is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 97/100, with no cautions. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

kde-mcp compared with similar skills

All 4 of these similar skills score higher than kde-mcp; compare them before choosing.

SkillScoreStarsUpdatedFormat
kde-mcp (this skill)by atassis761013d agoMCP Server
taste-skillby Leonxlnx100118.9k11d agoCLAUDE.md
Agent-Reachby Panniantong100113.3k12d agoCLAUDE.md
hyperframesby heygen-com10087.1ktodayCLAUDE.md
rufloby ruvnet10082.1ktodayCLAUDE.md

Frequently asked questions

How do I install kde-mcp?
Run claude mcp add atassis -- npx -y github:atassis/kde-mcp. The install tabs above show the steps for each supported agent.
Which AI agents does kde-mcp work with?
It is written for Claude Code and Claude Desktop, as a MCP Server file. Other agents that read the same format can often use it too.
Is kde-mcp safe to use?
It is MIT-licensed and scores 97/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is kde-mcp still maintained?
The repository was last updated 13 days ago, so kde-mcp is actively maintained.

kde-mcp

An MCP server that lets an AI agent observe and control a KDE Plasma 6 / Wayland desktop through programmatic primitives — the accessibility tree (AT-SPI), KWin Scripting, and libei via KWin's private EIS interface — with screenshots only as a fallback.

Status: working MVP. All planned tools are implemented; the server runs as a systemd-user daemon and is driven daily from Claude Code. Platform: KDE Plasma 6 on Wayland (KWin). Language: Rust (edition 2024). License: MIT.

It is the Linux/KDE counterpart to the computer-use capability Anthropic ships for macOS. Anthropic's version is macOS-only; this project targets KDE Plasma 6 on Wayland.

What makes it different

  • Accessibility-tree first; pixels as fallback. The primary action is act_on_element, which invokes an AT-SPI action on a referenced element — no coordinate hallucination, no input-emulation race, and roughly 5–30× cheaper than screenshot round-trips. Screenshots are an explicit escape hatch, not the default loop. (ADR-0005)
  • Response-as-instruction. Decision-relevant information is carried in structured response fields, not in tool-description prose — the agent reacts to a diff in the moment instead of a paragraph far back in context. (ADR-0004)
  • Native to KDE/Wayland, no portal dialogs. Input goes through org.kde.KWin.EIS.RemoteDesktop (libei), capture through org.kde.KWin.ScreenShot2, window control through org.kde.KWin.Scripting. No Xvfb, no Docker, no xdotool, no interactive portal prompts.
  • One mode per process. --mode=live drives the real desktop; --mode=virtual spawns a headless kwin_wayland --virtual session (optionally watchable over VNC) for isolated automation. No runtime mode-swapping, so there is no dual-context misrouting to reason about. (ADR-0003)
  • A deliberately small, safety-shaped tool surface — 12 tools. Tools consolidate along intent and split along safety, so the host's readOnlyHint/destructiveHint annotations do the right thing. Four tools that merely duplicated host-native capabilities (bash, text_editor.*, wait) were removed once they failed that test. (ADR-0006, ADR-0013)
  • A safety architecture adapted from Anthropic's design. Single-session mutation lease, per-app approval via KDE notifications, app-category tiering (view-only / click-only / full), terminal-excluded screenshots, and a Meta+Esc global cancel grabbed at the compositor via KGlobalAccel. Every destructive tool routes through a central policy gate. (ADR-0007)

For the full side-by-side with Anthropic's macOS computer-use, see docs/comparison-with-anthropic-computer-use.md. A broader, honest comparison against other Linux desktop-control MCP projects is planned (see docs/research/prior-art.md for the current survey).

Architecture

  • Three transports in one binary: stdio, Streamable-HTTP, and a Unix domain socket.
  • Singleton daemon. The daemon runs as a systemd-user service with socket activation and idle-stop, so multiple agent sessions share one process (one tray icon, one AT-SPI subscription) instead of contending. The preferred transport is a Unix domain socket with 0600 permissions; an HTTP transport on 127.0.0.1:7575 is the current default only because some MCP clients can't yet speak unix:// (there, the loopback bind is the trust boundary). (ADR-0011, ADR-0012, ADR-0015)
  • Audited dispatch. Every tool call writes one structured audit record; untrusted external strings are carried with a three-level trust tag (trusted / agent-authored / external) so the agent can tell apart what it authored from what the screen returned. (ADR-0008)

Install

See packaging/README-install.md for dependencies (Arch / Fedora / Debian), the build, the systemd-user units, and the Claude Code / Claude Desktop MCP configuration. In short:

cargo install --path .
./packaging/install.sh        # installs the .desktop authorization + systemd-user units

Then point your MCP client at http://127.0.0.1:7575/mcp.

Repository layout

src/                  Rust implementation (tools, transports, policy gate, AT-SPI / KWin / EIS plumbing)
examples/             Standalone EIS probers — minimal POCs for the libei input path
docs/specs/           The design specification (the load-bearing document)
docs/adr/             Architecture Decision Records (0001–0015) — one decision per file
docs/research/        Snapshot research that informed the design (Wayland/KDE primitives, prior art, LLM tool-design)
docs/plans/           Historical implementation plans (planning artifacts; the code is the source of truth)
docs/runbook/         A machine-executable verification protocol and product scenarios
tools/                Latency analyzer for the per-phase tool-call profiling
packaging/            .desktop authorization, systemd-user units, policy.toml example

Development

Contributors and AI coding agents: read CONTRIBUTING.md and AGENTS.md before opening a PR. The short version: Conventional Commits, cargo fmt + cargo clippy --workspace --all-targets -- -D warnings + cargo test --workspace must be clean before every commit, no unwrap()/expect() on user-reachable paths, and one ADR per non-obvious decision.

License

MIT — see LICENSE.

Related Skills

View on GitHub
GitHub Stars10
CategoryDevelopment
Updated13d ago
Forks2

Languages

Rust

Trust signals

97/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 info