Drive a real, visible browser that is already logged in. CDP for Chromium, WebDriver BiDi for Firefox.
Gaze is a browser automation tool with 22 tools covering navigation, element interaction, and session management. Critical gaps: (1) Most tool descriptions are extremely brief (5-15 chars), far below the 50-200 char baseline for LLM optimization. E.g., 'goto' = 'Navigate to a URL' (18 chars), 'map' = 'Extract interactive elements from the page' (41 chars). (2) Input schemas are present but minimal, most parameters lack detailed constraints, ranges, or format guidance. E.g., 'url' param has no validation hints; 'selector' params lack CSS selector format guidance. (3) No output schemas documented anywhere, LLMs cannot predict what these tools return. (4) No error handling guidance, tools like 'eval' and 'upload' have high risk but no recovery hints. (5) Parameter descriptions are generic; e.g., 'filter' in 'map' says 'Optional filter for element types' but does not enumerate valid types. (6) No idempotency or side-effect declarations despite many WRITE operations. (7) Tool composition is weak, 'login' and 'session' both manage auth but lack clear distinction. Average per-tool score: 38/100.
Clear an input field
Click an element on the page
Access browser console logs
Double-click an element
Download a file from the page
Evaluate JavaScript in the page context
Fill an input field with text
Navigate to a URL
Tool descriptions are critically short (5-41 chars), far below the 50-200 char LLM-optimized baseline. LLMs cannot determine when to select tools or what they return. E.g., 'goto' = 'Navigate to a URL', 'press' = 'Press a keyboard key'.
No output schemas documented for any tool. LLMs cannot predict return types, forcing them to guess what fields are available and risking failed downstream tool chains.
Parameter descriptions lack format guidance and constraints. E.g., 'selector' params do not specify CSS selector syntax; 'filter' in 'map' does not enumerate valid element types; 'key' in 'press' does not list valid key names.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | 2026-07-28+ | v2 |
Grant standing approval for write operations
Hover over an element
Load saved authentication session
Extract interactive elements from the page
Access network request logs
Press a keyboard key
Record browser activity to video
Revoke standing approval
Extract text content from page elements
Scroll the page
Select an option from a dropdown
Manage saved authentication sessions
Extract all visible text from the page
Upload a file to an input element
No error handling guidance. High-risk tools like 'eval' (arbitrary JavaScript execution), 'upload' (file I/O), and 'download' (file I/O) lack recovery hints or validation error messages.
Tool composition ambiguity: 'login' and 'session' both manage authentication but lack clear distinction. 'session' accepts a 'command' param (list/load/save/delete) but does not document which commands are safe vs destructive.