$ guiport▌

Why guiport.

Agents shouldn't drive desktop apps by guessing pixels. guiport exposes the desktop as structured data — accessibility tree first, screenshots and OCR only as fallback — so flows are deterministic, replayable, and fast enough to use as test infra.

Install.

Available on macOS, Windows, and Linux. macOS 13+ runs the full accessibility-selector stack; Windows & Linux ship standalone binaries today — screenshot, keyboard/mouse input, and app control — with a11y-tree selectors landing next.

brew install edihasaj/guiport/guiport
guiport doctor --fix   # grant Accessibility + Screen Recording
guiport doctor         # all green when ready

curl -fsSL https://raw.githubusercontent.com/edihasaj/guiport/main/scripts/install.sh | sh

# grab the latest standalone binary from Releases
# Windows: guiport-<ver>-windows-x64.zip  (runtime DLLs bundled)
# Linux:   guiport-<ver>-linux-x86_64.tar.gz (static stdlib, glibc 2.35+)

git clone https://github.com/edihasaj/guiport.git
cd guiport
swift build -c release
sudo cp .build/release/guiport /usr/local/bin/guiport

Quick start.

List running apps.

$ guiport apps --with-windows
Calculator      com.apple.calculator   pid=12345  windows=1
Finder          com.apple.finder       pid=614    windows=2
Code            com.microsoft.VSCode   pid=48047  windows=1

Inspect an app's accessibility tree.

$ guiport tree --app Calculator --pretty | head -20
{
  "id": "/AXWindow[0]",
  "role": "AXWindow",
  "name": "Calculator",
  "children": [
    {
      "role": "AXButton",
      "identifier": "Five",
      "bounds": {"x": 605, "y": 442, "width": 48, "height": 48},
      "actions": ["AXPress"]
    }, ...
  ]
}

Click by stable selector — not coordinates.

$ guiport click --app Calculator 'AXButton[identifier="Five"]'
{"path":"ax","selector":"AXButton[identifier=\"Five\"]","detail":"5"}

App doesn't expose structure? Visual fallback is automatic.

$ guiport click --app Figma 'AXButton[name="Save"]'
{"path":"ocr","detail":"Save @ 980,124"}

Record → Replay.

$ guiport record smoke.yaml --app Calculator
recording for Calculator — interact with the app, Ctrl+C to stop
recording saved → smoke.yaml

$ guiport run smoke.yaml
{"passed":true, "steps":[ ... ]}

CLI reference.

Full help: guiport <command> --help.

Selector grammar.

role[attr="value"][attr~="substring"][index]

Examples:

# Exact match by AX role + name
button[name="Save"]

# Substring match
AXButton[name~="Open"]

# Stable identifier (best — survives layout changes)
*[identifier="save_btn"]

# Pick the third match
AXMenuItem[name~="File"][2]

Attributes: role, name (title), value, identifier, description, text (any of name/value/description), index.

Role aliases: button → AXButton, checkbox → AXCheckBox, textfield → AXTextField, etc.

YAML test format.

name: calculator 5+3=8
app: Calculator
timeout_ms: 4000
steps:
  - find:   'AXButton[identifier="AllClear"]'
  - click:  'AXButton[identifier="AllClear"]'
  - wait:   80
  - click:  'AXButton[identifier="Five"]'
  - click:  'AXButton[identifier="Add"]'
  - click:  'AXButton[identifier="Three"]'
  - click:  'AXButton[identifier="Equals"]'
  - assert:
      find:   'AXStaticText[value~="8"]'
      exists: true

Step types: wait · activate · find · click · press · type · hotkey · screenshot · find_text · click_text · click_at · assert (existence + frontmost / front_title_contains / focused).

Package these flows as reusable, app-specific plugins in ~/.guiport/plugins — guiport plugin run <name> <action> — with no app knowledge baked into core.

Failures snapshot the AX tree (JSON) + screenshot + action log into artifacts/fail-* for fast diagnosis.

MCP server.

Plug guiport into an MCP-aware agent (Claude Code, opencode, others). One JSON object per line over stdio:

$ guiport serve --mcp
{"jsonrpc":"2.0","id":1,"method":"initialize"}
<= {"jsonrpc":"2.0","result":{"protocolVersion":"2024-11-05",...}}

{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{
   "name":"click","arguments":{
     "app":"Calculator",
     "selector":"AXButton[identifier=\"Five\"]",
     "fallback":"ocr"
   }
}}

Tools exposed: doctor, apps, observe, tree, find, click (+ fallback), click_at, find_text, click_text, type, hotkey, screenshot, run, plugin_list, plugin_run.

Platforms.

One CLI, one YAML schema, one MCP tool list across every platform — available on macOS, Windows, and Linux. macOS runs the full accessibility-selector stack today; Windows and Linux ship standalone binaries with screenshot, keyboard/mouse input, and app control, and the a11y-tree selectors slot in behind the same contract.

One contract, every platform.

            guiport
       CLI · MCP · runner
              │
   ┌──────────┼──────────┐
   ▼          ▼          ▼
 macOS      Windows    Linux
 shipped     available  available

Same selectors, same YAML, same MCP tools — across every platform. macOS runs the full stack today; Windows and Linux ship binaries now with capture + input, and the a11y selectors slot in behind the same shared contract. Your replay tests never know which OS is underneath.