Loading ToolBench
Server has 7 tools with complete input schemas and basic descriptions. However, descriptions are minimal (10-50 chars), lacking context on WHEN to use each tool or what happens on failure. Parameters lack type constraints (enums, ranges, patterns). No output schemas documented. Error handling is minimal, no recovery guidance. Tool names follow verb_noun convention but descriptions don't explain dependencies (e.g., puppeteer_navigate must be called before puppeteer_click). Security concern: launchOptions parameter accepts arbitrary browser args, creating injection risk despite allowDangerous flag. No pagination, no batch operations, no idempotency guarantees.
Click an element on the page
Execute JavaScript in the browser console
Fill out an input field
Hover an element on the page
Navigate to a URL
Take a screenshot of the current page or a specific element
Select an element on the page with Select tag
Descriptions are too brief (10-50 chars) and lack context. 'Click an element on the page' doesn't explain when to use puppeteer_click vs puppeteer_evaluate, or what happens if the selector doesn't match. LLMs cannot infer prerequisites or error recovery.
No output schemas documented. Callers don't know what fields to expect from puppeteer_screenshot (is it a base64 string? a URL? a file path?). This forces LLMs to guess and breaks tool chaining.
launchOptions parameter accepts arbitrary browser arguments without strong validation. Despite allowDangerous flag, merging environment config with user input creates injection risk. Dangerous args list is hardcoded but incomplete (e.g., --disable-gpu, --disable-dev-shm-usage not listed).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | 2026-07-28+ | v2 |
No error handling guidance. If puppeteer_click fails (selector not found, element not clickable), the error response doesn't tell the LLM what to do next. Should suggest: try puppeteer_screenshot to debug, or use puppeteer_evaluate to inspect the DOM.
Global state (browser, page, consoleLogs, screenshots) is not thread-safe and persists across calls. If two agents call puppeteer_navigate concurrently, they share the same browser instance, causing race conditions and cross-contamination.