Loading ToolBench
A Windows/Linux MCP server that exposes practical workstation tools for local agents.
This server has 28 tools with significant quality gaps. While most tools have descriptions and input schemas are documented in the spec, the descriptions are often generic or incomplete, parameter descriptions frequently lack detail, and output schemas are not formally documented. The tool set is unfocused, it spans C compiler tooling, web SEO auditing, image optimization, database extraction, and shell command execution, suggesting multiple unrelated domains bundled into one server. Many tools lack actionable error guidance. The descriptions tend toward imperative summaries rather than LLM-optimized context (e.g., 'Execute shell commands' is bare; 'Execute arbitrary shell commands with output capture and timeout management' would be better). Per-tool scores average to 42, placing this in the 'poor' range.
Audit local HTML files or one URL for common SEO problems.
Check C source file syntax without compilation.
Compile C source files to an executable using available C compilers.
Detect available C compiler toolchains and tools on the system.
Safely evaluate a C math expression.
Extract data from MySQL databases using read-only queries.
Extract data from PostgreSQL databases using read-only queries.
Critical: IRREVERSIBLE tools (run_shell, run_cmd, run_powershell) lack error handling guidance and do not implement confirmation/dry-run pattern. An LLM can execute destructive commands like 'rm -rf /' with no safety gate or recovery path.
Critical: Tool descriptions lack LLM-optimized context. Most are imperative summaries (15-30 chars) that do not explain WHEN to use a tool or what its output enables. E.g., 'Detect available C compiler toolchains and tools on the system' does not explain that this is a discovery tool to call before compile_c_program, nor what fields are in the response.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Extract data from SQLite databases using read-only queries.
Format C source code using available formatters.
Generate C build configuration files (CMakeLists.txt, Makefile).
Generate a C header file with fixed-point arithmetic functions.
Generate a C header file with lookup table generation.
Generate a C header file with utility math functions.
Generate a C header file with matrix algorithm functions.
Generate a C header file with matrix math functions.
Generate C test harness code for numeric computations.
Generate a C header file with polynomial evaluation functions.
Generate a C header file with statistical functions.
Generate a C header file with vector math functions.
Inspect C source code structure including includes, defines, and functions.
List all available tools and their categories.
Create smaller image files while preserving the main visual result by default.
Preprocess C source files with macro expansion and includes.
Execute commands using cmd.exe or shell.
Execute PowerShell scripts.
Execute shell commands.
Scan C source code for common memory safety issues.
Scan C source code for numeric computation risks including overflow and underflow.
High: Output schemas are not documented. No visible JSON schema for response structures. An LLM cannot plan the next tool call or extract the right field if it does not know what fields the response contains. E.g., detect_c_toolchains returns a dict but the LLM must infer field names.
High: Parameter descriptions are incomplete or missing for many tools. E.g., run_shell has 'max_output_chars' but no explanation of what happens when output exceeds the limit (is it truncated? does the tool fail?). compile_c_program has 'preferred_compiler' but does not explain the fallback order if not found.
High: The server bundles 28 unrelated tools (C compiler suite, SEO audit, image optimization, database extraction, shell execution). This violates the single-responsibility principle and makes the tool set difficult for LLMs to navigate. Consider splitting into domain-specific servers: allcanuse-c-mcp, allcanuse-seo-mcp, allcanuse-db-mcp, allcanuse-shell-mcp.
Medium: Security concern: run_shell, run_cmd, and run_powershell accept arbitrary command strings with no validation. An LLM could be prompt-injected into executing 'rm -rf /'. No mention of sandboxing, restrictions, or audit logging.
Medium: Database extraction tools (extract_sqlite_content, extract_postgresql_content, extract_mysql_content) claim to use 'read-only queries' but have no validation to enforce SELECT-only. An LLM could pass DROP TABLE or INSERT statements.
Medium: compile_c_program accepts 'run_after_compile' and 'run_args' parameters, effectively converting it into an execution tool. This conflates compilation and execution into a single tool, making it harder for LLMs to reason about side effects and harder to audit when compilation succeeded but execution failed.
Medium: Error messages are not documented. No guidance on what errors compile_c_program, run_shell, or audit_seo might return, or how to recover. An LLM will not know if a 'compiler not found' error is retryable or permanent.
Low: Tool naming is generic in places. 'evaluate_c_math_expression' does not clarify what 'safely' means or what constraints apply to expressions. 'optimize_images_for_memory' mixes 'memory' (RAM) with 'file size', which is it?