Loading ToolBench
Unified interaction server combining browser automation, Windows UI automation, and vision/OCR capabilities. Provides tools for web scraping, desktop element inspection, screenshot capture, and OCR.
AI-Hands exhibits severe definition quality gaps across 37 tools. Most tools (18/37) lack visible input schemas in the provided source, only hands_app_action shows a complete schema. Descriptions vary wildly: some are detailed (hands_app_action, find_and_click), but 19 tools have minimal or generic descriptions (uia_get_state, uia_list_windows, vision_screenshot, etc. show only 1-3 word descriptions). Parameter documentation is sparse; most tools lack per-parameter descriptions. Many tools appear inferred rather than explicitly registered (uia_*, vision_*, browser_* tools show no schema definitions in source). Deprecated tools (find_and_click, read_screen_text, window_screenshot) are still registered, creating confusion. No visible error handling guidance, output schemas, or recovery patterns. This is a STDIO-only server with 37 loosely-defined tools, typical of early-stage automation frameworks, not production-grade MCP servers.
Attach to an existing browser instance with optional stealth mode
Click an element in the browser with optional accessibility reference
Focus an element in the browser with optional accessibility reference
Get intercepted/logged network requests captured by routes. Credential values, URL query values, fragments, and request/response body values are always redacted; body shape is preserved when structured.
Hover over an element in the browser with optional accessibility reference
Launch a browser instance with optional stealth mode for anti-detection
Scroll in the browser with optional accessibility reference
Select an option in a browser element with optional accessibility reference
18 of 37 tools have no visible input schema in source code. Only hands_app_action shows a complete JSON Schema definition. All uia_*, vision_*, and most browser_* tools lack schema documentation, making it impossible for LLMs to understand parameter types and constraints.
19 tools have minimal descriptions (1-3 words: 'Get...', 'List...', 'Perform...'). Minimal descriptions prevent LLMs from selecting the right tool.
Three tools marked DEPRECATED (find_and_click, read_screen_text, window_screenshot) remain registered and callable. Deprecated tools create confusion and waste LLM reasoning cycles deciding between old and new variants. Remove or clearly gate them.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 31 | <=2025-11-25 | v2 |
Stop trace recording and save sanitized entries to a new JSON file. Existing files are never overwritten. URL query values, fragments, click text, script source, and resolved parameter values are always redacted.
Stop trace recording and return sanitized entries. URL query values, fragments, click text, script source, and resolved parameter values are always redacted.
Type text into a browser element with optional accessibility reference
Type text into a browser element with optional accessibility reference
DEPRECATED: use hands_click (offset_x/offset_y subsume the find_and_click offset behavior; rungs 6+7 cover OCR + UIA fallback). Find text on screen via OCR, then click its location via UIA. Combines vision + UIA in one call. When window_title is provided, focuses that window first. When text isn't found on the visible screen, automatically tries focusing other windows to find it.
Window management meta-tool. Actions: open, close, focus, minimize, maximize, restore, snap_left, snap_right, snap_top, snap_bottom. Combines UIA window operations with WindowMatch for flexible window targeting, save dialog detection, post-action verification, and monitor stickiness.
Central multi-monitor scope tool — one ability front door for monitor-aware operations
DEPRECATED: use hands_capture (accepts optional window_title to focus a named window before OCR) or vision_screenshot_ocr for screen-only OCR. Take a screenshot and return all text via OCR. Optionally target a specific window by title.
Click a UI element on the desktop
Find a UI element on the desktop by various criteria
Focus a window on the desktop
Get the current state of UI elements on the desktop via Windows UI Automation
Press a key or key combination on the desktop
List all open windows on the desktop
Poll for UI automation events on the desktop
Read the value of a UI element on the desktop
Scroll a UI element on the desktop
Execute a keyboard shortcut on the desktop
Type text into a UI element on the desktop
Watch for changes in a UI element on the desktop
Analyze an image for visual content
Compare two images and return differences
Find a template image within a screenshot
Load an image from disk
Perform OCR on a screenshot to extract text
Take a screenshot of the screen or a specific window
Take a screenshot and perform OCR in one call
Poll the screen until specific text appears (via OCR) or a template image is found. Returns when found or times out.
PARTIALLY DEPRECATED: for default (behind=false) capture, prefer `vision_screenshot` (focus the window yourself if needed, then screenshot). For obscured/background windows, prefer the new `vision_screenshot_hidden_window` tool (PrintWindow API, always-on). This combined tool remains registered for backward compatibility. Focus a window by title and take a screenshot.
No visible error handling guidance, recovery patterns, or output schemas in any tool definition. LLMs cannot know what to do on failure or what fields to expect in responses. No actionable error messages documented.
Parameter descriptions are missing or generic across most tools. E.g., browser_click accepts 'a11y_ref' but lacks guidance on format, where to find it, or what happens if invalid.