Stealth browser MCP server for AI agents — drives a real Firefox via Camoufox, passes Cloudflare, clicks like a human.
Cybr Ghost provides 13 well-named browser automation tools with clear, action-verb naming (navigate, click, type_text, screenshot). All tools have descriptions (10 - 200 chars, within baseline 34 - 392 range). Input schemas are present and typed for all tools. However, descriptions lack LLM-optimized depth, most are 1 - 2 sentences without WHEN/WHY context, dependency hints, or recovery guidance. Parameter descriptions are minimal (e.g., 'Snapshot index of element to click' lacks format/constraint detail). Output schemas are not formally documented, responses are inferred as text/string. Error handling is absent: no guidance on retryability, no actionable error messages, no recovery paths. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels (READ_ONLY vs WRITE). Composition is sound, tools chain well (navigate → snapshot → click), but no batch variants or pagination for future scaling.
Click the element with the given snapshot index [N].
Execute arbitrary JS in page context. Escape hatch for things the other tools can't do. Example: `document.querySelectorAll('.price').length` Return value is stringified.
Get outer HTML of a selector (default: body). Truncated to max_chars.
Get all visible page text (document.body.innerText). Useful for reading articles / extracting content.
Navigate to a URL. wait_until: 'domcontentloaded' | 'load' | 'networkidle'. Returns a snapshot of interactive elements on the loaded page.
Open a new tab (optionally navigate). Subsequent calls target the new tab.
Press a keyboard key globally (e.g. 'Enter', 'Escape', 'Tab', 'ArrowDown', 'Control+a').
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite explicit risk levels. Tools marked WRITE (click, type_text, press_key, eval_js, new_tab) should declare destructiveHint=true; READ_ONLY tools should declare readOnlyHint=true. This prevents agents from accidentally treating side-effecting calls as safe to retry.
Descriptions lack LLM-optimized depth. Most are 1 - 2 sentences without WHEN/WHY context, dependency hints, or recovery guidance. E.g., 'Re-snapshot the current page. Use after dynamic content loads or modal opens.' should explain: 'Call after dynamic content loads (e.g., after clicking a button that triggers AJAX) or when a modal opens. Returns updated element list with new indices.' This guides agent planning.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Capture a PNG screenshot. Use when snapshot() isn't enough and you need to SEE the page.
Scroll. direction: 'up'|'down'|'top'|'bottom'. amount: pixels for up/down.
Re-snapshot the current page. Use after dynamic content loads or modal opens.
Type text into input element [N]. submit=True presses Enter after. clear=True empties field first.
Wait N seconds (capped at 30) for dynamic content, then return fresh snapshot.
Wait for a CSS/XPath selector to appear, then snapshot.
No error handling or recovery guidance. Tools lack actionable error messages. E.g., if click(index=999) fails (index out of bounds), the agent receives a raw exception instead of 'Index 999 out of range (0 - 42 available). Call snapshot() to refresh element list.' This forces agents to guess recovery steps.
Output schemas not formally documented. Responses are inferred as text/string. E.g., navigate() returns a snapshot string, but the structure (URL, Title, Elements list with [index] tag, href, text, etc.) is not declared in the tool definition. LLMs cannot plan downstream calls without knowing what fields to extract.
Parameter descriptions lack format/constraint detail. E.g., 'wait_until' in navigate() says 'Wait condition: domcontentloaded | load | networkidle' but should be an enum constraint in the schema, not a description. 'index' in click() lacks bounds (0 - MAX_ELEMENTS). 'key' in press_key() lacks examples (Enter, Escape, Control+a). Agents cannot validate inputs without explicit constraints.