Loading ToolBench
MCP server that executes shell commands on Windows and returns intelligently minified, context-aware output. Understands pytest, git, npm/pnpm/yarn, and generic commands. Returns only errors, warnings, summaries and key results — never raw logs.
Single tool with adequate schema and reasonable description, but lacks critical parameter descriptions, output schema documentation, and error handling guidance. The tool name 'run_command_minified' is somewhat generic and doesn't follow verb_noun convention clearly. Input schema is present with types, but parameter descriptions are minimal. No documented output structure, no error recovery guidance, and no indication of which errors are retryable. The description (158 chars) is within acceptable range but lacks WHEN to use it and prerequisites.
Execute a shell command on Windows and return an intelligently minified, context-aware output. Understands pytest, git, npm/pnpm/yarn, and generic commands. Returns only errors, warnings, summaries and key results — never raw logs.
Output schema not documented. Tool returns JSON with 'status', 'command_type', 'errors', 'warnings', 'summary', 'key_data' fields, but LLMs have no formal specification of these fields' types, ranges, or when each is populated. This forces LLMs to infer structure from examples.
Parameter descriptions are minimal. 'command' lacks format/constraint guidance (e.g., 'Windows cmd.exe or PowerShell syntax'). 'timeout' description is only 34 chars and doesn't explain consequences of timeout or valid range beyond default.
No error handling guidance. Tool can fail (command not found, timeout, permission denied, etc.) but returns no recovery hints. Per pattern:recovery-guide, errors must tell the LLM what to do next. Current code returns generic JSON error dict with no actionable next steps.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Tool name 'run_command_minified' is vague about scope. Does it run ANY command, or only safe ones? Does 'minified' mean output is truncated, summarized, or filtered? Ambiguous names cause LLMs to guess wrong. Clearer name: 'execute_shell_command_with_summary' or 'run_windows_command_filtered'.
No security/permission documentation. Tool executes arbitrary shell commands with WRITE risk. No mention of what commands are blocked, what permissions are required, or audit trail. Per pattern:permission-gate and pattern:audit-trail, destructive tools must declare scope and log calls.