MCP server that exposes CadQuery script execution in a containerized sandbox with stdio transport
Single tool with partial schema and French-language description. Input schema present but minimal (only 'script' string param, no constraints). Output schema documented in docstring but not in JSON Schema format. Description is 194 chars and explains the tool's purpose, but lacks guidance on when to use it, prerequisites, or error recovery. No parameter-level descriptions beyond the tool docstring. Tool name 'run_cadquery_script' is clear and verb-based, but the implementation shows this is a specialized domain tool with high risk (WRITE, Docker execution) that lacks proper error categorization and recovery guidance.
Execute un script CadQuery dans un conteneur isolé. Le script doit écrire son solide dans /work/out.step. Retourne produced, error_message, artifact_path, size_bytes, returncode, timed_out. produced=false avec error_message=null est le signal d'arrêt de la boucle de reprise.
Input schema lacks parameter-level descriptions and constraints. The 'script' parameter has no description field in the JSON Schema, violating pattern:tool-description. LLMs cannot infer that this expects CadQuery Python code or understand format/length constraints.
Output schema is documented only in the docstring (free text), not as a formal JSON Schema. The tool returns a dict with 6 fields (produced, error_message, artifact_path, size_bytes, returncode, timed_out) but this structure is not machine-readable in the MCP schema. Agents cannot validate or plan downstream calls reliably.
Error handling does not categorize errors or guide recovery. ToolError is raised for 'docker not found', 'timeout', and 'missing report.json', but the LLM receives no guidance on whether to retry, ask the user, or escalate. Pattern:recovery-guide requires actionable error messages.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 45 | 2025-06-18+ | v2 |
Tool description is in French ('Execute un script CadQuery...'), making it inaccessible to English-speaking LLMs and violating the principle that descriptions must be prompt-engineered for clarity. MCP servers should use English for maximum compatibility.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite the tool performing a WRITE operation (Docker execution, file creation). The tool is marked Risk: WRITE but this is not reflected in the MCP schema, leaving the LLM unaware of the operation's irreversibility.