Kira

   ██╗  ██╗██╗██████╗  █████╗
   ██║ ██╔╝██║██╔══██╗██╔══██╗
   █████╔╝ ██║██████╔╝███████║
   ██╔═██╗ ██║██╔══██╗██╔══██║
   ██║  ██╗██║██║  ██║██║  ██║
   ╚═╝  ╚═╝╚═╝╚═╝  ╚═╝╚═╝  ╚═╝
        local · MCP · computer-use

kira-mcp is a local Model Context Protocol server that gives any MCP-compatible agent host (Claude Desktop, Claude Code, Cursor, Cline, Continue, …) full computer-use capabilities on the host machine.

https://github.com/user-attachments/assets/b9b65987-32b1-43f5-bdf0-006ea71c6d13

Kira solving a CAPTCHA end-to-end — one perceive_screen, click-ready pixels, no human in the loop.

Built and tuned for Windows. macOS and Linux are best-effort — most tools work, but some UI conventions differ.

• Vision — perceive_screen is the agent's one-shot "look at the screen" tool. It grabs the current display in memory, runs the local microsoft/OmniParser-v2 YOLO icon-detector on it, and returns an annotated image plus JSON with each element's {id, bbox, cx, cy, confidence} in absolute screen pixels — so the agent can pipe cx, cy straight into mouse_click. No API key, no network call.
• Desktop automation — pixel-accurate mouse control, keyboard input (incl. chords and key holds), and clipboard read/write via pyautogui, mss, and pyperclip.

The server speaks stdio JSON-RPC and is launched as a child process by your agent host.

Quick start

You need Python 3.10+ and an MCP host (the examples use Claude Desktop). Pick your OS.

pip install kira-mcp

{
  "mcpServers": {
    "kira-mcp": { "command": "kira-mcp" }
  }
}

pip install kira-mcp

{
  "mcpServers": {
    "kira-mcp": { "command": "kira-mcp" }
  }
}

sudo apt install python3-tk python3-dev scrot xdotool   # or your distro's equivalent
pip install kira-mcp

💡 Want text matching too? pip install "kira-mcp[ocr]" enables find_element(text="Send") so the agent can click by label instead of eyeballing icons.

Install options

The Quick start above uses pip. For an isolated global install that won't touch any project's site-packages:

pipx install kira-mcp

Either form installs the kira-mcp console script and registers every tool module.

The OmniParser-v2 YOLO icon-detector weights (icon_detect/model.pt, ~39 MB) ship inside the wheel — no separate download step. They are loaded and warmed up from disk at server startup.

Working on kira-mcp itself? Clone and install editable: git clone https://github.com/Anmol202005/kira-mcp.git && cd kira-mcp && pip install -e .

If your clone is missing the weights (e.g. a shallow checkout, or you stripped them), restore them with:

hf download microsoft/OmniParser-v2.0 \
  icon_detect/model.pt \
  icon_detect/model.yaml \
  --local-dir src/kira_mcp/weights

Or point to a model.pt elsewhere on disk by setting KIRA_YOLO_WEIGHTS in your environment.

Other agent hosts

The Quick start covers Claude Desktop. Every other host uses the same config shape — point it at the kira-mcp command:

• Claude Code — claude mcp add kira-mcp -- kira-mcp
• Cursor / Cline / Continue / Windsurf — add the same mcpServers block to the host's MCP config.

Prefer not to rely on the installed script? Use the module form anywhere a command is expected:

{ "command": "python", "args": ["-m", "kira_mcp"] }

After editing config, restart the host. All tools appear under the kira-mcp namespace.

Tools

Vision

Mouse

Keyboard

Key names accept any value from pyautogui.KEYBOARD_KEYS, plus common aliases (ctrl, alt, shift, cmd/command, win/windows/super, meta, esc/escape, enter/return, space/spacebar, pgup/pageup, pgdn/pagedown, del, ins).

Screen

Clipboard

Typical agent loop

perceive_screen()                             # → annotated JPEG inline + JSON: {width, height, elements: [{id, bbox, cx, cy, confidence}, …]}
# agent picks an element by id, reads its (cx, cy) — already in absolute screen pixels
mouse_click(x=cx, y=cy)
perceive_screen()                             # verify the action landed

One tool to look, one tool to act, repeat until done.

Layout

src/kira_mcp/
├── __main__.py        # entry — `python -m kira_mcp` or `kira-mcp`
├── _mcp.py            # shared FastMCP instance + system instructions
├── lib/
│   ├── keys.py        # key-name normalization for pyautogui
│   ├── cache.py       # screenshot + per-element OCR hash cache
│   └── ocr.py         # optional OCR backend shim
└── tools/
    ├── __init__.py    # side-effect imports → registers tools
    ├── parse.py       # `perceive_screen` — screenshot + local YOLO icon-detector
    ├── screen.py      # `screen_size` + the `Region` model
    ├── mouse.py
    ├── keyboard.py
    └── clipboard.py

Add a new tool by writing a function decorated with @mcp.tool() (imported from kira_mcp._mcp) and importing the module from tools/__init__.py.

Local development

pip install -e .
python -m kira_mcp        # stdio server — drive it from your MCP host

Safety

pyautogui.FAILSAFE is enabled at startup — slamming the mouse to the top-left corner raises FailSafeException and aborts whatever the agent was doing. Leave it on. The server explicitly does not expose a way to disable it from tool calls.

Wayland note

On Linux Wayland sessions, raw X11 screen grabs return a black buffer. GNOME Wayland additionally blocks programmatic screenshots from unprivileged callers. If perceive_screen returns a black image (or no detections at all), log in to an X11 session, or switch to a Wayland compositor that ships wlr-screencopy (Hyprland, Sway, river, Niri) or a KDE Plasma session.

License

MIT — see LICENSE.