Loading ToolBench
Model Context Protocol (MCP) server for local system statistics, CPU, RAM, Disk, GPU, and process monitoring.
Strong foundation with 10 well-named tools following verb_noun convention (get_*, search_*). All tools have detailed descriptions (150-350 chars, well above 10-char minimum). Input schemas present with types and enums. However, output schemas are not explicitly documented in source code, descriptions mention return fields but no formal schema definitions visible. Error handling guidance is minimal. Tool descriptions are comprehensive but could be more concise for LLM token efficiency. No tool annotations (readOnlyHint, destructiveHint) despite all being read-only operations.
Retrieve comprehensive CPU performance metrics and thermal telemetry. Returns processor model, logical cores, physical cores, current/min/max clock speeds (GHz), total and breakdown load (user, system, idle percentages), per-core utilization percentages, main and max package temperatures (°C), thermal health status ('NORMAL' <75°C, 'HOT' 75-89°C, 'CRITICAL' >=90°C), temperature sensor source, and individual per-core temperatures when available. Use to diagnose thermal throttling or compute bottlenecks.
Retrieve storage capacity and free space metrics for a specified drive path or mount point. Returns total capacity (GB), free space (GB), used space (GB), and utilization percentage. Use to identify low-disk-space warnings (>85% utilized) and verify available storage for software installations or large downloads.
Analyze per-process graphics memory (VRAM) utilization and evaluate GPU memory headroom. Returns GPU model, total VRAM (MB/GB), used VRAM, free VRAM, remaining VRAM margin (GB), overall VRAM usage percentage, and a ranked list of active processes using the GPU with PID, application name, dedicated VRAM (MB/GB), and shared VRAM (MB). Use to identify memory-heavy AI models (e.g. Ollama, PyTorch, Stable Diffusion, LM Studio), 3D renderers, or games, and determine if sufficient VRAM remains to load new models without out-of-memory errors.
Retrieve global graphics card (GPU) telemetry and attached display configurations. Returns GPU controller specifications, vendor, model, total VRAM (MB), real-time GPU compute utilization percentage, GPU core temperature (°C), PCIe bus interface, and connected display details (resolutions, refresh rates in Hz, connection types, and primary monitor flag). Use to monitor graphics hardware health and multi-monitor setups.
Output schemas not formally documented. Descriptions mention return fields (e.g., 'Returns GPU model, total VRAM, used VRAM...') but no explicit JSON Schema definitions visible in source. LLMs cannot reliably parse unstructured field lists.
No tool annotations despite all 10 tools being read-only. Adding readOnlyHint=true to all tools would signal safety to agents and enable optimizations (e.g., safe to retry, no side effects).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Retrieve precise physical RAM utilization metrics. Returns total system RAM, free available RAM, used RAM, and memory utilization percentage in either GB or MB. Use this to analyze memory saturation, check headroom for large workloads, or assess RAM pressure.
Inspect local network interface cards (NICs), IP routing addresses, and network topology. Groups interfaces into active external network adapters (Ethernet, Wi-Fi, virtual bridges) and internal loopback interfaces, returning IPv4/IPv6 addresses, subnet families, hardware MAC addresses, and scope flags. Use to determine host IP addresses, network interface configuration, and local connectivity.
Perform real-time network throughput and bandwidth telemetry by sampling transfer rates over a 1-second interval. Returns per-interface download and upload transfer speeds in both Kilobytes/sec (KB/s) and Megabits/sec (Mbps), cumulative received and transmitted data in Megabytes (MB), and identifies the busiest network interface. Use to diagnose network congestion, active bandwidth hogging, or measure current transfer speeds.
Query background system services, daemon processes, and Windows services. Returns total service counts, count of currently running services, count of stopped services, and a detailed list with internal service name, friendly display name, execution status ('Running' vs 'Stopped'), startup configuration mode ('Auto', 'Manual', 'Disabled'), and associated PID when running. Supports filtering by execution status and keyword search across service names and descriptions. Use to inspect service health, identify stopped essential services, or audit background daemons.
Identify and rank resource-intensive applications and processes currently running on the local computer. Can sort by real-time CPU consumption percentage or physical RAM memory usage (MB), group multiple instances of the same application (e.g. multi-process browsers, Discord, Electron apps) into unified entries with instance counts and PID arrays, or return individual granular PID processes. Use to diagnose system slowdowns, high CPU spikes, and memory leaks.
Search and inspect running processes by name or substring query (e.g. 'chrome', 'node', 'docker', 'python', 'ollama'). Returns whether matching processes are actively running, total count of active instances, aggregated CPU utilization percentage, aggregated physical RAM memory usage (MB), and a detailed breakdown of each individual instance with PID, process name, individual CPU %, and memory (MB). Use to verify if an application is active, hanging, or consuming excess resources.
Error handling guidance absent. Tools return data or fail silently. No recovery hints (e.g., 'If GPU not detected, check drivers' or 'If process not found, try search_process with partial name'). Agents cannot self-correct on failures.
Descriptions are verbose (150-350 chars) and include implementation details ('Returns GPU model, total VRAM (MB/GB), used VRAM, free VRAM, remaining VRAM margin (GB)...'). Baseline for A+ tools is 50-200 chars. Trim to essential intent and use parameter constraints for format details.
Prompts reference tools that do not exist in the tool registry (e.g., 'get_system_health', 'get_system_overview', 'get_hardware_specs', 'get_usb_devices'). Prompts will fail at runtime when agents try to call non-existent tools.