Loading ToolBench
AI分镜脚本与视频生成智能体，遵循 MCP (Model Context Protocol) 规范。支持生成专业分镜、视频、镜头重生成、配额查询等能力。
ShotFlow defines 7 tools with complete JSON schemas and descriptions in Chinese. Naming follows verb_noun convention (generate_*, check_*, list_*, regenerate_*). Schemas are well-structured with enums and type constraints. However, descriptions lack English translations (barrier for non-Chinese LLMs), parameter descriptions are minimal in some tools (check_quota, check_health have empty properties), and output schemas are entirely undocumented. Error handling guidance is absent. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Composition is sound, tools chain logically (generate_storyboard → generate_video → check_video_status).
检查ShotFlow服务运行状态、版本号、可用功能列表。
查询当前用户的免费配额剩余次数或Pro会员状态。
查询视频生成任务的当前状态。任务完成后可获取视频文件URL。
从剧本/创意描述生成专业分镜脚本。支持6种创作类型和7种视觉风格，输出包含场景-镜头层级结构、台词、声音设计、情绪标注、导演备注和AI图像提示词。这是ShotFlow的核心能力。
从分镜镜头描述或AI提示词生成短视频片段。支持多种分辨率，返回任务ID用于后续状态查询。需要Pro会员权限。
列出最近的视频生成任务及状态。
对分镜脚本中的单个镜头重新生成AI描述和提示词。可修改镜头类型、情绪方向等参数，其他镜头保持不变。
Output schemas completely undocumented. No tool returns a documented response structure. LLMs cannot plan downstream calls or extract required fields (e.g., what does generate_storyboard return? Is it a JSON object with 'scenes', 'shots', 'aiPrompt'?). This forces agents to guess field names and risks context loss.
Descriptions in Chinese only. Non-Chinese-speaking LLMs (e.g., Claude, GPT-4 in English mode) will not understand tool purpose, when to call them, or parameter semantics. Descriptions must be in English or include English translations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2025-06-18+ | v2 |
check_quota and check_health have empty inputSchema.properties (no parameters documented). Even though they accept no parameters, the schema should be explicit: properties: {} with description 'No parameters required.' Current form is ambiguous.
No error handling guidance. Tools like generate_video (requires Pro membership) and generate_storyboard (calls external AI APIs) will fail in predictable ways (quota exceeded, API timeout, auth failure). Descriptions do not explain what to do on failure or how to recover.
No tool annotations despite clear risk levels. generate_video and regenerate_shot are WRITE operations; check_* and list_* are READ_ONLY. Annotations (readOnlyHint, destructiveHint, idempotentHint) are missing, preventing agents from reasoning about side effects and retry safety.