The server defines 11 Selenium automation tools with complete JSON schemas and basic descriptions. However, definitions fall short of production quality in multiple critical areas: (1) Descriptions are generic and lack actionable context for LLM selection (e.g., 'Finds an element on the page' vs. 'Use to locate interactive elements before performing actions like clicking or sending keys'); (2) Error handling returns generic strings ('Error finding element: ...') rather than actionable recovery guidance; (3) No parameter validation constraints (enums for locator strategies, min/max for timeout); (4) Missing output schema documentation, tools return unstructured strings instead of structured objects with typed fields; (5) No indication of which parameters are mutually exclusive or dependent; (6) No batch operations despite patterns (e.g., click_element could benefit from a click_elements accepting arrays). Schema quality is moderate (all tools have input schemas with types and descriptions), but lack of structured output, minimal error guidance, and generic descriptions prevent higher scores.
Clicks an element on the page.
Close the current browser session
Performs a double click on an element.
Drags an element and drops it onto another element.
Finds an element on the page.
Gets the text of an element.
Hovers over an element.
No structured output schemas documented. All tools return unstructured string responses (e.g., 'Element found using id="xyz"') instead of typed JSON objects. LLMs cannot reliably parse responses or chain to downstream tools.
Descriptions are too generic and lack selectivity context. 'Finds an element on the page' (45 chars) does not explain when to use find_element vs. click_element vs. send_keys. Baseline for A-grade: 50-200 chars with explicit selectivity guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 56 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 45 | - | v1 |
Navigates the browser to a specified URL.
Performs a right click (context click) on an element.
Sends keys to an element (typing).
Start a browser (supports Chrome and Firefox)
No parameter constraints as enums. The 'by' parameter accepts arbitrary strings but only 6 values are valid (id, css, xpath, name, tag, class). Undeclared constraint forces runtime validation and invites hallucinated values from LLMs.
Error messages are non-actionable and do not guide recovery. Returns bare exception strings ('Error finding element: NoSuchElementException'). Should return 'Element not found with xpath="..." (timeout 10s). Verify the element exists and the locator is correct.'
No permission checks or audit logging. The server manages browser sessions in process state (state['drivers']) with no isolation, no user tracking, no operation audit trail. Shared state across agents/users risks session hijacking and makes incident investigation impossible.
No idempotency or confirmation for destructive operations. close_session is IRREVERSIBLE and modifies state (state['drivers'] delete). No dry-run mode, no confirmation step. Agents can accidentally close shared sessions.
Missing support for multiple simultaneous sessions. The design uses a single global 'current_session' variable. If an LLM wants to interact with two browsers in parallel, it cannot, switching session_id is required after each tool call. Better design: accept optional session_id parameter to all tools.
Parameter descriptions are too terse. E.g., 'The value for the locator strategy' (41 chars). Does not clarify examples or format constraints (e.g., xpath vs. css selector syntax differences). Baseline: 50-150 chars with examples and patterns.