MCP bridge that connects AI agents (Cursor, Claude Code, VS Code Copilot, etc.) to the AIPex browser extension. Auto-spawns a shared daemon to support multiple simultaneous clients.
The AIPex MCP Bridge defines 15 tools with complete JSON schemas and explicit descriptions. However, quality is uneven. Tools fall into two categories: tab management (tools 1-7) with minimal descriptions (20-50 chars), and browser interaction (tools 8-15) with detailed, multi-sentence descriptions. Parameter descriptions are present across most tools but vary significantly in specificity. Error handling guidance is largely absent. The 'computer' tool (15) stands out with exceptionally detailed documentation (~800 chars), while tab tools like 'get_all_tabs' (1) have bare-minimum descriptions. No tool annotations (readOnlyHint/destructiveHint) are present despite clear WRITE vs READ_ONLY classifications in the code. Output schemas are not documented in the definitions provided. Overall, the server demonstrates functional tool design but lacks the polish expected of production-grade agent integrations.
Click an element using its unique UID from a snapshot
Close a specific tab or the current tab
[HIGH-COST FALLBACK] Coordinate-based mouse/keyboard interaction using screenshot pixels. PREFER UID-BASED TOOLS FIRST: For clicking buttons, filling forms, or hovering elements, use search_elements to get UIDs, then use click/fill_element_by_uid/hover_element_by_uid. These are faster and more reliable. USE THIS TOOL ONLY WHEN: - search_elements returned 0 matches after trying 2 different query patterns - UID-based actions failed twice (element not interactable) - The goal requires visual/pixel-level interaction: canvas apps, drag-and-drop, sliders, charts, hover-only menus. PREREQUISITE: If you choose coordinate actions, you MUST first call capture_screenshot(sendToLLM=true). Coordinates are in screenshot pixel space. * Click element centers, not edges. Adjust if clicks miss.
Create a new tab with the specified URL
Fill an input element using its unique UID from a snapshot
Tab management tools (get_all_tabs, get_current_tab, ungroup_tabs) have minimal descriptions (20-30 chars), far below the 50-200 char production baseline. These descriptions do not explain WHEN to use each tool or differentiate overlapping functionality.
No tool annotations present despite clear risk categorization. Code marks tools as READ_ONLY or WRITE, but tool definitions lack readOnlyHint, destructiveHint, or idempotentHint attributes. This prevents agents from understanding which operations are safe to retry.
Output schemas are not documented in tool definitions. No response structure is specified for any tool. LLMs cannot predict what fields to expect or plan downstream tool chains. Production baseline: 100% of A-tier tools document output schemas.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 68 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Fill multiple form elements at once using their UIDs from a snapshot
Get all open tabs across all windows with their IDs, titles, and URLs
Get information about the currently active tab
Get the complete content from a code editor (Monaco, CodeMirror, ACE) or textarea without truncation. Use this before filling to avoid data loss.
Get detailed information about a specific tab
Hover over an element using its unique UID from a snapshot
[FAST - USE FIRST] Search for elements in the current page using glob/grep patterns against the DOM snapshot. Returns matching elements with their UIDs for direct UID-based interaction. GLOB SYNTAX: * matches any characters (e.g. button* finds all buttons), ? matches exactly one character, [abc] matches any of those characters, {a,b,c} matches any of those alternatives (e.g. {button,input}* finds all buttons and inputs), Patterns are case-sensitive by default; use [Ll] to match both cases. WORKFLOW: 1. Call search_elements with a broad pattern to discover elements 2. Elements in the result have uid= attributes (e.g. uid=btn-42) 3. Pass that UID to click(tabId, uid) or fill_element_by_uid(tabId, uid, value) 4. If 0 results after 2 different patterns, fall back to capture_screenshot(sendToLLM=true). This is the PREFERRED first step — much faster and cheaper than screenshots.
Switch to a specific tab by ID or URL pattern
Remove all tab groups in the current window
Upload a file to a file input element (<input type="file">) on the page using a local file path. WORKFLOW: 1. Provide the tabId and an absolute local file_path 2. If the page has multiple file inputs, use input_index to select which one (0 = first) 3. Optionally provide uid from a snapshot if you know the exact element. NOTE: Most websites hide the actual <input type="file"> behind a styled button. This tool handles both visible and hidden file inputs automatically. AFTER UPLOAD: take a screenshot to verify the file was accepted, then proceed to submit the form.
Error handling guidance is absent. No tool description explains what errors might occur, how to recover, or what the agent should do next. For example, 'click' does not explain what happens if the UID is invalid or the element is not interactable.
Parameter descriptions vary in quality. Many lack format constraints, ranges, or usage guidance. For example, 'query' in search_elements is described as a 'Search query string with grep/glob pattern support', but the actual glob syntax documentation is in the tool description, not the parameter annotation where the LLM expects it.
Tool naming could be clearer: 'computer' is vague and does not convey intent from the name alone. Industry convention would be 'interact_by_coordinates', 'pixel_click', or 'coordinate_action'. The name 'computer' requires reading the description to understand it's a fallback coordinate-based interaction tool.
Optional parameters are not clearly marked. For example, 'switch_to_tab' has required=[] even though both 'tabId' and 'urlPattern' are optional alternatives. This forces the LLM to guess whether it must provide one, both, or either. Schema should clarify mutual dependency: 'either tabId OR urlPattern required, not both'.
No pagination or result limits documented for 'get_all_tabs' or 'search_elements'. If a browser has hundreds of open tabs or a page has thousands of DOM elements, the agent receives unbounded results that exhaust context. Production pattern: always include limit and offset/cursor parameters.