Taimuraz Kaitmazov

← work

kde-mcp — computer-use for the Linux desktop

An MCP server that lets an agent actually drive a Linux (KDE Plasma 6 / Wayland) desktop — the kind of capability Anthropic ships for macOS and Windows, still rare on Linux. Built from scratch, solo, in Rust.

The bar before me

Computer-use tooling is mature on macOS and Windows and thin on Linux — and what exists tends to lean on pixel-level screenshots and coordinate clicks, which are brittle and unsafe (a moved window or a re-theme breaks them). Wayland's security model makes naive screen-scraping harder still.

What I built

kde-mcp drives applications through the accessibility tree — the same structured interface a screen reader uses — rather than by clicking pixels. The agent acts on named, typed elements, so actions are more reliable and far easier to reason about and gate. A safety policy layer sits in front of destructive actions, and the design decisions are written down as ADRs rather than living in my head.

Rust
language
KDE Plasma 6 / Wayland
target
a11y tree
drives apps via
policy gate · ADR
safety

Verify

It's open source — read the code, the ADRs, and the safety policy:

Public repository — the accessibility-tree approach, the safety gate, and the ADRs are all in the open.

github.com/atassis/kde-mcp