Loading ToolBench
Multi-model vision MCP server (GLM-4.6V, Kimi, Qwen-VL, GPT-4o) that lets text-only coding models in Claude Code see and understand images.
Two tools with reasonable descriptions and parameter schemas, but significant gaps in output documentation and error handling guidance. vision_describe_image has a detailed description (Chinese, ~400 chars) and well-structured input schema with enums and defaults. vision_retry_last_image lacks input parameters entirely (empty schema) and minimal description (~150 chars). Neither tool documents return types or error recovery paths. Parameter descriptions are present but could be more actionable for LLM reasoning.
使用视觉模型识别图片并返回文本描述。支持任意 OpenAI 兼容的视觉模型（火山方舟豆包、Kimi、智谱 GLM-4V、通义 Qwen-VL、OpenAI GPT-4o 等）。 支持输入本地图片路径或 http(s):// 图片 URL。MCP 内部会自动将图片转换为 base64 后调用视觉模型 API（多数视觉模型不支持直接传 URL）。 Args: - image (string, 必填): 图片路径或 URL。 - 本地路径示例: "C:/Users/user/screenshot.png" 或 "/tmp/img.jpg" - URL 示例: "https://example.com/img.png" - prompt (string, 可选): 额外识图指令，会附加到当前 mode 的提示词模板后；不传则仅使用模式模板（默认 auto 自动分类）。最长 4000 字符 - mode (string, 可选): auto / design_rebuild / prototype_understanding / bug_screenshot / general，默认 auto Returns: 文本内容。默认会先自动判断图片类型，再输出适合设计图还原、原型图理解或 bug 截图分析的结构化结果，并附带 token 使用情况元信息。
当用户要求重新识别之前失败的图片（如说"重试"/"重新识别"）时调用。无参数，重试最近一张失败的图片。 代理会从暂存的失败图原图重新调用视觉模型识别：成功则返回描述，失败则返回错误原因。 若返回"没有可重试的失败图"或端点不可达，说明失败图已过期（代理可能重启过）或暂存被淘汰，请让用户重新粘贴图片。
vision_retry_last_image has empty input schema (no parameters defined). Tool is stateless but lacks any documentation of return type or error conditions.
No output schemas documented for either tool. LLMs cannot infer what fields to expect in responses, forcing unstructured parsing and increasing hallucination risk.
Error handling lacks recovery guidance. vision_retry_last_image description mentions 'no retry available' and 'endpoint unreachable' but does not tell LLM what action to take next (ask user, try alternative, etc.).
vision_describe_image description is in Chinese. While technically complete, LLMs trained primarily on English may misinterpret nuanced instructions. Consider English translation or bilingual descriptions.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | <=2025-11-25 | v2 |
Parameter 'prompt' has a default value ('请详细描述这张图片的内容。') that is language-specific and may not align with LLM's native language. Defaults should be language-agnostic or omitted.