easy-uiauto

English | 简体中文

Turn a Windows desktop interface into inspectable, reusable UI commands.

easy-uiauto combines a Python automation library, an MCP server, action recording, and an Obsidian-compatible UI knowledge vault. It can locate controls through Windows UI Automation, learn the useful parts of a visible application, and expose verified operations as commands such as main.keypad.enter-digit-6.click.

The project is useful for Windows test automation, RPA, agent tooling, and building reusable operation knowledge for desktop software. It is Windows-only and currently released as a 0.x project.

⚠️ Visual scanning and workflow exploration are still evolving. Review learned commands before using them for messages, payments, deletion, uploads, account changes, or other operations with external effects.

✨ What It Includes

• 🖱️ Desktop automation: click, double-click, right-click, drag, scroll, keyboard input, hotkeys, window activation, and text entry.
• 🎬 Action recording: record real operations, highlight the selected control, and generate reusable Python actions with a complete LOCATION.
• 🔌 MCP server: use the same automation surface from Codex, Claude Code, or another MCP client.
• 🔎 Visual-first learning: identify the important regions and controls in one screenshot, then map them back to stable local UIA controls without walking the whole tree by default.
• 🧠 State-aware semantics: distinguish stable functions from dynamic values, retain names and avatars, connect related controls, and explain visible states such as a Send button disabled while its message input is empty.
• 🧭 Semantic UI CLI: turn verified controls into searchable application commands instead of rediscovering coordinates for every task.
• 🖼️ Control overlays: number every learned or executable control in one click-through overlay, save annotated page images, and preview operation targets.
• 🧪 Effect learning: compare before/after screenshots, inspect changed UIA properties and new windows, and store an evidence-backed success condition.
• 🗂️ Readable knowledge: keep pages, controls, images, interactions, and quarantine records as Markdown/YAML/PNG files that work directly with Obsidian and version control.
• 📐 Stable records: persist structured LOCATION identities and normalized visual hints, never process IDs, window handles, or absolute desktop rectangles.
• 🛟 Layered lookup: unique AutomationId, contextual XPath, stable property combinations, saved image states, local OCR, then opt-in remote vision.
• ⚡ Fast anchors: use a unique window-scoped AutomationId before XPath, while keeping hierarchy as the fallback for missing or duplicate IDs.

🚦 Capability Status

Area	Status	Notes
Python mouse, keyboard, window, and UIA APIs	Stable	The original library surface
Recording and replay	Stable	Produces reusable structured actions
MCP tools and client setup	Beta	Codex and Claude Code installers included
Visual-first scan and generated UI CLI	Beta	Requires an OpenAI-compatible vision endpoint
Before/after effect learning	Beta	Stores local evidence and learned success conditions
Recursive workflow exploration	Experimental	Restricted to reversible commands and bounded depth

Automated workflow exploration intentionally excludes scrolling and dragging. The base automation library still provides those operations for explicit scripts and direct MCP calls.

📦 Installation

Requirements: Windows and Python 3.11 or newer.

Install from PyPI:

pip install easy-uiauto

Or install from source:

git clone https://github.com/Poggi-Tang/easyautomation.git
cd easyautomation
pip install -e .

🔌 MCP Setup

Install the optional MCP dependencies when using easy-uiauto from an MCP client:

pip install "easy-uiauto[mcp]"

Install local OCR and image-template fallback support when needed:

pip install "easy-uiauto[mcp,vision]"

The vision extra provides OpenCV template matching and the Python Tesseract adapter. OCR also requires the system Tesseract executable and relevant language data (for example eng or chi_sim). The MCP tools are find_control_by_image, click_by_image, find_text_on_screen, and click_text_on_screen.

For ordinary Python code, install only the local vision extra and call the rectangle segmenter directly. It does not start MCP, call OCR, or upload data:

pip install "easy-uiauto[vision]"

from easy_uiauto.visual import segment_screen_rect

result = segment_screen_rect(
    {"left": 100, "top": 200, "right": 900, "bottom": 800},
    min_width=6,
    min_height=6,
    save_annotated=True,
)
for box in result["boxes"]:
    print(box["rect"], box["parent_id"])
print(result["relations"], result["groups"])

For a coarse UIA canvas whose internal elements have no UIA nodes, pass the control result or its bounds directly to the local rectangle detector:

get_control_at_position(x, y)
detect_visual_elements(rect=<the returned control object>)

detect_visual_elements captures only the supplied rectangle and returns numbered class-agnostic boxes, absolute screen rectangles, input-relative normalized rectangles, containment relationships, confidence, timing, and an annotated preview path. Its local layout graph adds right-of, below, alignment, overlap, and adaptive-gap relations, then builds variable-size union groups. Group signatures intentionally exclude element width, height, and observed content, so changing name lengths and multiline content remain the same structure. Repeated vertical items and table rows remain separate. The tool does not traverse UIA, run OCR, infer semantics, download a model, or call a remote API.

Remote visual location does not run a local model or require a vendor SDK. Set an OpenAI-compatible vision endpoint and credentials, then use find_control_by_vision or click_by_vision:

EASY_UIAUTO_VISION_API_URL=https://your-api.example/v1/chat/completions
EASY_UIAUTO_VISION_API_KEY=your-api-key
EASY_UIAUTO_VISION_MODEL=your-vision-model

Those tools upload the current screenshot to the configured endpoint only for that request. Use them as a final fallback after UIA, OCR, or image matching.

The MCP server ships with the library and reuses the same automation APIs:

easy_uiauto --help
easy_uiauto --version
easy_uiauto

The setup commands register, inspect, or remove the MCP entry through the client's own CLI. Existing entries are not overwritten by the basic installer.

For a complete Codex installation:

pip install --upgrade "easy-uiauto[mcp,vision]"
easy_uiauto --full-setup-codex \
  --vision-url https://your-api.example/v1/chat/completions \
  --vision-model your-vision-model

Full setup asks for a missing API key through hidden input, installs the Python vision dependencies and Tesseract when needed, replaces only the easy_uiauto Codex entry, installs the two bundled skills, and checks UIA, OCR, and the remote vision endpoint. Validation uses generated images rather than the current desktop. Vision settings are read dynamically by a current-version MCP process. Restart Codex once after installing or updating the package, MCP entry, or skills; changing only the vision settings does not require a second restart.

For a minimal Codex deployment with remote AI vision, use the quick setup command. When the API key is not already present in the Windows user environment, it prompts once through hidden terminal input or a password dialog for non-interactive agents. It persists the three vision variables, replaces only the easy_uiauto Codex MCP entry, and skips OCR installation and UI tests:

easy_uiauto --quick-setup-codex \
  --vision-url https://your-api.example/v1/chat/completions \
  --vision-model your-vision-model

You can also give Codex this deployment instruction:

Install the latest easy-uiauto[mcp,vision] from PyPI, then run easy_uiauto --full-setup-codex --vision-url URL --vision-model MODEL; do not inspect unrelated projects or search the web; report every validation result and elapsed time, then ask me to restart Codex once. Do not request another restart only because the vision environment variables changed.

easy_uiauto --install-codex
easy_uiauto --show-codex-config
easy_uiauto --uninstall-codex

easy_uiauto --install-claude-code
easy_uiauto --show-claude-code-config
easy_uiauto --uninstall-claude-code

Codex registration uses its global config.toml. Claude Code uses the user scope. Restart the client after installing, updating, or removing the server.

The long-running TCP service is also available:

easy_uiauto_service --help
python -m easy_uiauto.mcp.service --port 9876

🧭 Learn an Application

The package includes two Codex skills:

• $easy-uiauto-learning scans and explores visible Windows applications.
• $easy-uiauto-operate runs commands from an existing application vault.

After setup and one Codex restart, a natural-language request is enough:

Use $easy-uiauto-learning to learn the currently open WeChat window and build
its control knowledge and UI commands.

🔎 Light Scan

A light scan observes one visible page. It does not click the application.

easy_uiauto_ui scan "Window title"
easy_uiauto_ui apps
easy_uiauto_ui commands <app-id>
easy_uiauto_ui show <app-id>

The default visual-first strategy sends one target-window screenshot to the configured vision endpoint, identifies useful regions and controls, and maps those coordinates back to local UIA controls. It stores stable LOCATION records, control crops, meaning, aliases, risk, and generated commands. Secondary monitors are supported. The scan also saves <page-id>.annotated.png and briefly draws the same numbered controls over the live window. Use --no-overlay when visual feedback is not wanted.

MCP scans report their task ID, stage, elapsed time, and control-verification count every five seconds. If a client stops waiting, call get_ui_learning_status before retrying. A running task continues in the MCP process and must not be started twice. Plain learning requests stop after this scan. Interactive explore_ui_workflows runs only when deep or autonomous exploration is explicitly requested.

Use --strategy full-uia for diagnostics when visual targeting is not suitable. This mode walks the visible UIA tree and performs a separate semantic pass, so it is slower and produces more records.

🧪 Deep Exploration

Deep exploration starts with a light scan, executes known reversible commands, and records what changed. New pages and dialogs can be scanned recursively.

Use $easy-uiauto-learning to deeply learn the currently open WeChat window with
the safe policy, at most 30 actions, and recursion depth 3. Do not send messages,
delete data, log out, or run any confirmation-required operation.

easy_uiauto_ui explore <app-id> --policy safe --max-actions 30 --max-depth 3
easy_uiauto_ui interactions <app-id>

For each operation, the explorer saves before/after screenshots, local pixel differences, changed UIA properties, new or transient windows, the observed success condition, and the recovery result. It deduplicates page/command states and stops when recovery fails or another user/window interrupts the run.

safe executes only commands classified as safe. supervised may also execute reversible state-changing commands. External, destructive, and confirmation-required commands are never explored automatically. Scrolling and dragging are not implemented in the explorer.

▶️ Run Learned Commands

easy_uiauto_ui search <app-id> "search terms"
easy_uiauto_ui show <app-id> --include executable
easy_uiauto_ui run <app-id> <page.region.control.action> --text "optional text"
easy_uiauto_ui batch <app-id> '["main.keypad.enter-digit-6.click", "main.keypad.enter-addition-operation.click"]'
easy_uiauto_ui learn-effect <app-id> <command> --recover
easy_uiauto_ui teach <app-id> <control-id> "Meaning" intent "Description"

Runtime lookup follows unique AutomationId → contextual XPath → Name/ClassName/ControlType combination → saved image states → local OCR → opt-in remote vision. Enable the last step with --allow-vision-fallback or allow_vision_fallback=true. Stale, missing, or ambiguous controls are moved to quarantine instead of being clicked. Single and batch commands preview their resolved targets in red by default; pass --no-highlight or highlight=false to remove the roughly 100 ms visual-preview wait.

set-text first verifies UIA value assignment, then falls back to clipboard paste for Unicode text. It reports success only after an accessible value or exact clipboard readback confirms the write; a local image change is diagnostic evidence only. The user's previous text clipboard value is restored after pasting.

Use a batch only for commands on the same stable page. Split the sequence after navigation. Commands with external or destructive effects require --confirm or confirm=true.

🗂️ Knowledge Vault

The default vault is ~/easy_uiauto_vault. Set EASY_UIAUTO_KNOWLEDGE_DIR to use another location. Markdown/YAML and PNG files are the source of truth:

applications/<app-id>/
├── pages/                 page records
├── regions/               functional regions
├── controls/              verified and observed controls
├── interactions/          before/after effect records
├── images/                page, control, and interaction images
├── quarantine/            stale or ambiguous controls
└── operations/UI-CLI.md   generated command catalog

.easy_uiauto/index.json is a disposable search cache. Rebuild it with easy_uiauto_ui reindex <app-id> after editing Markdown by hand.

Control and region geometry is stored only as a window-relative normalized_rect with geometry_role: visual-hint-only. It is useful for annotations and visual comparison, but never identifies or clicks a control. Live overlays and operations resolve the current LOCATION first, then use a unique image template or local OCR fallback. Version 0.6 automatically removes legacy absolute rectangles, PIDs, and window handles when an older vault index is rebuilt.

🔐 Privacy and Safety

• Light scanning sends the selected target-window screenshot to the configured endpoint.
• Effect learning sends target-window before/after images and crops of related new windows. Full desktop snapshots remain local in the vault.
• API keys are read from environment variables and are not written into the knowledge vault.
• PIDs, native window handles, and absolute desktop coordinates are runtime observations only and are not persisted as application knowledge.
• Inferred control meanings remain separate from locator verification and observed operation effects. Use teach to correct a meaning; it cannot bypass failed location checks.
• Existing reliable controls are retained when a visual-first rescan omits them.

The same workflow is exposed through get_ui_learning_readiness, get_ui_learning_status, scan_window_knowledge, show_ui_controls, list_ui_knowledge_apps, search_ui_knowledge, search_ui_knowledge_batch, list_ui_commands, run_ui_command, run_ui_commands, learn_ui_command_effect, explore_ui_workflows, list_ui_interactions, teach_ui_control, and rebuild_ui_knowledge_index.

--full-setup-codex installs or updates both bundled skills. Use --install-codex-skills when only the skills need to be refreshed.

For MCP client configuration, start the server with python -m easy_uiauto.mcp.server. Control-vector persistence is optional. To enable it, set EASY_UIAUTO_CONTROL_VECTOR_DB_DIR to a directory containing control_vector_store.py; otherwise capture tools still return records but do not persist them.

Control lookup uses the library's canonical LOCATION object rather than a flat selector. Obtain it from a recorded action or from get_control_at_position, then pass the returned LOCATION object directly to find_control:

{
  "WindowName": "My Application",
  "Name": "Save",
  "ClassName": "ButtonClass",
  "ControlType": "ButtonControl",
  "foundIndex": 1,
  "AutomationId": "saveButton",
  "Xpath": [
    {"ControlType": "WindowControl", "Name": "My Application", "searchDepth": 1},
    {"ControlType": "ButtonControl", "Name": "Save", "foundIndex": 1, "searchDepth": 2}
  ],
  "Img": "",
  "PARAMETERS": {}
}

find_control also accepts a complete recorded action containing LOCATION and the complete result from get_control_at_position. Legacy flat arguments remain supported. A present AutomationId is tried before the stored XPath and ignores dynamic Name, foundIndex, and recorded geometry; XPath handles missing or duplicate IDs.

🧰 Python API Examples

Basic Control Operations

from easy_uiauto.ctrl import Controller

# Left click on a control
Controller.left_click(
    ActionTitle="Click OK Button",
    WindowName="My Application",
    Name="OK",
    ClassName=None,
    ControlType="ButtonControl",
    foundIndex=0,
    AutomationId="",
    Xpath=[],
    Img="",
    PARAMETERS={}
)

# Input text into a field
Controller.input_text(
    ActionTitle="Enter Username",
    WindowName="Login Dialog",
    Name="Username",
    ClassName=None,
    ControlType="EditControl",
    foundIndex=0,
    AutomationId="",
    Xpath=[],
    Img="",
    PARAMETERS={"输入文本": "test_user"}
)

# Keyboard shortcut
Controller.key_group(
    ActionTitle="Save File",
    WindowName="Notepad",
    Name="",
    ClassName=None,
    ControlType="",
    foundIndex=0,
    AutomationId="",
    Xpath=[],
    Img="",
    PARAMETERS={"组合键": "ctrl+s"}
)

Recording User Actions

from easy_uiauto.record import run_record

# Start recording user actions
run_record(write_file=True)
# Press ESC to stop recording
# Generated script will be saved to Record{timestamp}.py

🏗️ Project Structure

easyautomation
├── .github/
│   └── workflows/
│       ├── ci.yml
│       ├── publish.yml
│       └── release.yml
├── src/
│   └── easy_uiauto/
│       ├── __init__.py
│       ├── ctrl.py          # Core controller (mouse/keyboard actions)
│       ├── draw.py          # Visual feedback (control highlighting)
│       ├── record.py        # Action recording
│       ├── utils.py         # Control location and utilities
│       ├── mcp/             # MCP server, scanner, vault, and UI CLI
│       └── skills/          # Bundled Codex learning and operation skills
├── tests/
├── CHANGELOG.md
├── LICENSE
├── README.md
├── README.zh-CN.md
└── pyproject.toml

🚀 Release Automation

The repository uses GitHub Actions for validation and publishing:

• CI tests Python 3.11, 3.12, and 3.13, then builds the package.
• Tagged releases build and publish through PyPI Trusted Publishing.
• Build artifacts include both an sdist and a wheel.

🛠️ Development

pip install -e .[dev]
pytest
ruff check .

📚 More Examples

For more examples, please refer to the test files in the demo/ directory or check the docstrings in the source code.

📄 License

MIT License. See LICENSE.

💬 Contact

Scan the QR code to add me on WeChat: