FastMCP server providing system information tools: GPU info, kernel info, and top CPU processes
Three tools with minimal but present descriptions and no input validation. gpu_summary and get_kernel have zero parameters and trivial schemas. top_processes has a single integer parameter with type and description, but lacks output schema documentation. No error handling guidance, no parameter constraints (e.g., limit bounds), and descriptions are under 100 chars, below the 194-char production baseline. Tools are read-only and well-named (verb_noun pattern), but lack the structured output documentation and recovery guidance expected of production tools.
Return full kernel string as in `uname -a`.
Return one‑line GPU summary (`nvidia-smi -L`). Gracefully degrades if nvidia‑smi is absent.
Return a table of the *limit* most CPU‑intensive processes (default 10).
No output schema documentation. Tools return plain strings; LLMs cannot infer structure or plan downstream calls. Production tools document return types explicitly.
top_processes lacks parameter constraints. 'limit' has no min/max bounds; LLMs could pass negative or absurdly large values. Should specify 1 - 1000 range.
No error handling guidance. If nvidia-smi or uname fails, tools return raw subprocess errors or empty strings. LLMs need actionable recovery hints (e.g., 'Try running on a Linux system with NVIDIA drivers').
Descriptions are 34 - 65 chars, below the 194-char production baseline. gpu_summary and get_kernel lack context on WHEN to use them or what they return. Descriptions should answer: What? When? What does it return?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
top_processes output is unstructured plain text. Production tools return structured JSON with typed fields (pid, command, cpu_percent, mem_percent). Unstructured output forces LLMs to parse text, wasting tokens and risking errors.