nekoro-browser exposes 7 tools with mixed quality. Naming is action-oriented (capture_screenshot, upload_file, close_tabs) but some tools are overly generic (cdp, exec_python). Descriptions exist but are brief (avg ~80 chars) and lack WHEN/WHY guidance. Schemas are present for all tools with proper JSON types, but parameter descriptions are minimal. Two tools (cdp, exec_python) are escape hatches that bypass normal tool composition. Error handling is absent, no guidance on retryability, user-fixable errors, or recovery paths. Output schemas are undocumented. The server prioritizes raw CDP access over high-level abstractions, which is intentional but limits LLM reasoning.

• Document output schemas for all tools. E.g., capture_screenshot returns {"data": "base64-encoded-png", "width": int, "height": int}. page_info returns {"url": string, "title": string, "tab_id": int}. This enables LLMs to chain tools and extract required fields.
• Expand tool descriptions to 80-150 chars with WHEN/WHY guidance. E.g., 'Capture a screenshot of the current page as PNG. Use this to verify page state or extract visual information. Returns base64-encoded image data.'
• Add parameter descriptions for all inputs. E.g., 'tabs: Array of tab IDs (integers) to close. Obtain tab IDs from page_info() or new_tab(). Closing a tab is irreversible.'
• Implement error handling with recovery guidance. E.g., if upload_file fails with 'File not found', return 'File not found at {path}. Check the path and try again.' If close_tabs fails with 'Tab already closed', return 'Tab {tab_id} is already closed or does not exist.'
• Consider wrapping cdp and exec_python with higher-level abstractions. E.g., instead of raw cdp, offer navigate_to_url, click_element, fill_form. This improves LLM reasoning and reduces hallucination risk.