Loading ToolBench
MCP server for JuiceFS file system diagnostics and management, supporting both CSI driver troubleshooting and direct JuiceFS operations
This MCP server provides 13 tools for JuiceFS/Kubernetes diagnostics. All tools are defined with visible registrations in pkg/csi/tools.go and pkg/juicefs/tools.go using mcp-go. However, critical quality issues significantly limit the score: (1) Descriptions are entirely in Chinese, making them opaque to English-speaking LLMs and violating tool-description pattern norms for LLM optimization; (2) Parameter descriptions are sparse and lack format/constraint guidance; (3) No output schemas are documented anywhere in the provided source, forcing LLMs to guess what fields are returned; (4) Error handling and recovery guidance are not visible; (5) Tool names are generally clear and action-oriented (get_, find_, bench_, stats_, accesslog_), but lack sufficient parameter constraint documentation. The tool set is coherent for a Kubernetes diagnostic domain, but falls well short of production quality due to missing schema documentation and Chinese-only descriptions.
通过挂载点获取基于文件系统访问日志的实时监控数据
通过挂载点进行 juicefs 的性能测试
查看机器上的挂载点
获取对应节点上的 CSI Node Pod
获取排查 JuiceFS CSI 挂载问题的流程
获取应用 Pod 使用的 JuiceFS PV
根据 pv 获取对应节点上 Mount Pod 的日志
获取 Pod 日志
All tool descriptions are in Chinese (e.g., '获取排查 JuiceFS CSI 挂载问题的流程'), making them unusable for English-speaking LLMs and violating the tool-description pattern which requires descriptions optimized for LLM interpretation. LLMs cannot select tools or understand intent from non-English descriptions.
No output schemas are documented in the source code. Tools like get_pod, get_node, get_log_of_pod return complex Kubernetes objects, but no documentation shows LLMs what fields to expect. This violates the response-shaper pattern and forces LLMs to guess at output structure.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 37 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 42 | - | v1 |
通过挂载点查看客户端的载参数
根据 pv 获取对应节点上的 JuiceFS Mount Pod，可以查看 Mount Pod 的配置，包括 Mount Pod 的资源限制、挂载点、镜像等
根据 node 名获取 node yaml
根据 pod 名获取 pod 的 yaml，可以查看 pod 的所有信息，包括 pod 使用的 PVC、所在节点等
通过挂载点实时统计 JuiceFS 性能指标
Tools with no parameters (get_handle_flow, find_mountpoint) have empty input schemas ({}). While technically valid, the description for get_handle_flow is extremely brief (12 chars in Chinese) and provides zero guidance on what the tool returns or when to call it.
Parameter descriptions lack actionable guidance. For example, 'tailLines' is described only as '获取的日志行数' (number of log lines) with no default, range (e.g., 1-1000), or guidance on typical values. 'interval' in stats_in_juicefs has no units specified (seconds? milliseconds?).
No error handling or recovery guidance visible in the source. If a pod is not found, a node is unavailable, or a mountpoint does not exist, there is no documented error response or recovery strategy. Violates error-classification and recovery-guide patterns.