Loading ToolBench
Collection of utility scripts and hooks for CTF analysis, dynamic instrumentation, image generation, LLM benchmarking, and development workflow automation. Not a traditional MCP server.
This server exhibits severe definition quality gaps across nearly all dimensions. Tool names lack action verbs (pe_info, recon, notify, collect, preflight_edit, verify_edit, audit_log, branch_protection, commit_quality, large_file_warner, precompact_wrapup, benchmark). Descriptions are present but many are verbose (200-400+ chars) and lack clarity on WHEN to use each tool or what the LLM should do next. Input schemas are visible for most tools but lack proper type constraints and validation guidance. No output schemas are documented. Error handling is absent, tools provide no recovery guidance. Security concerns are severe: audit_log and branch_protection operate on sensitive Git operations without permission gates or audit trails. The notify tool accepts arbitrary file paths without sanitization. Overall, this reads as a collection of utility scripts hastily wrapped as MCP tools rather than a thoughtfully designed agent interface.
PostToolUse hook on Bash that audits all executed commands with automatic redaction of secrets (API keys, tokens, passwords). Logs to ~/.claude/bash-commands.log with timestamps.
LLM Benchmark Runner for Ollama. Runs inference benchmarks on specified model across multiple context sizes (2048, 4096, 8192, 16384, 32768) with questions covering logic, reasoning, coding, and debugging. Measures tokens/sec, TTFT, wall time, and GPU stats.
PreToolUse hook on Bash that blocks/warns operations on protected branches (main, master, production, release, prod). Prevents force pushes to protected branches and direct commits. Supports per-repo opt-out via ~/.claude/branch-protection-skip.txt.
Collects generated images from Codex with three-layer fallback strategy: main path (prompt-save), fallback 1 (generated_images directory), fallback 2 (rollout JSONL base64 decoding). Uses marker file for safety to prevent picking up images from previous executions.
PreToolUse hook on Bash (git commit) that blocks commits containing debugger statements and hardcoded secrets in staged files. Detects OpenAI/Anthropic API keys, GitHub tokens, AWS access keys, and Google API keys.
PreToolUse hook on Read that warns at 500KB and blocks reading files larger than 2MB. Suggests using offset/limit, Grep, or tail for large files to avoid context overflow.
Tool names lack action verbs. Names like 'pe_info', 'recon', 'notify', 'collect', 'audit_log' do not start with clear verbs (get_, create_, send_, analyze_, etc.). LLMs cannot infer intent from these names alone and must read full descriptions to disambiguate.
No output schemas documented. Tools like pe_info, recon, benchmark return complex structured data (JSON objects with nested fields) but provide no schema definition. LLMs cannot plan downstream tool calls or extract required fields without knowing the response structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
Channel-agnostic traffic-light notifier for unattended loops. Sends notifications via Telegram, Discord, Slack, or iMessage with optional media attachment and auto-compression of images.
PE static reconnaissance script - parses PE header without external tools, outputs static information needed for bypass procedures including architecture, ImageBase, entry point, ASLR, DEP, manifest, section analysis with entropy and packer detection, import analysis with attack surface hints, and suspicious strings (IP, URL, config paths, window titles)
PreCompact hook that reminds about uncommitted changes, unpushed commits, and recent desktop modifications before context compression. Non-blocking; provides systemMessage to alert user of potential unsaved work.
Pre-flight check for edit mode execution. Validates that Pillow is installed and measures base image dimensions for canvas locking in prompts. Runs before calling Codex to prevent generation of unverified images.
Frida passive reconnaissance script - first-time dynamic analysis instrumentation that observes target program behavior without modification. Hooks network (WS2_32), UI (CreateWindowEx, MessageBox), string comparison (lstrcmp, CompareString), crypto APIs, file operations, process creation, registry access, time verification, and memory scanning.
Verification tool for edit mode output. Validates whether Codex performed in-place modification or regenerated a similar-looking image. Supports two modes: bgremove (expects pixel-perfect preservation of subject with transparent background) and local (expects localized changes clustered in one area).
Security: audit_log and branch_protection operate on sensitive Git operations without permission gates, audit trails, or user confirmation. These tools can block/allow commits and log commands but provide no mechanism to verify the calling agent has authority to enforce such policies.
Security: notify tool accepts arbitrary file paths without sanitization. An LLM could be tricked into passing a path like '../../../etc/passwd' or a command injection payload. No path validation or sandboxing is evident.
Error handling absent. No tool provides recovery guidance, error classification, or actionable next steps. Tools fail silently or return raw errors without telling the LLM what to do next (retry, ask user, abort).