Loading ToolBench
IntelliSearch - A comprehensive MCP server providing multiple toolkits for browser automation, file operations, mathematical calculations, and system utilities
IntelliSearch has 17 tools with mixed quality. Naming is generally verb-first and clear (get_*, operate_*, search_*). However, critical gaps exist: descriptions range from adequate to minimal, parameter documentation is sparse with many parameters lacking descriptions, input schemas are present but incomplete (many parameters missing type definitions or descriptions), and output schemas are undocumented. Error handling is minimal. The server exhibits a 'thin wrapper' pattern with limited LLM-optimization. Most tools fall into the 40-55 range due to missing parameter descriptions and lack of output schema documentation. Tools like 'rm' (destructive) and 'take_screenshot' lack confirmation/dry-run patterns.
Append content to a file.
Safely evaluate a mathematical expression and return the result. Supported expressions and functions: abs, round, min, max, pow, sqrt, sin, cos, tan, log, pi, e
Read file content. Supports text files and PDF (extracts text). Excel preview is disabled.
Copy a file or directory (recursive).
Reads the text content from the local system clipboard.
Get the current date in the local timezone.
Get the current time in the local timezone.
Output schemas completely undocumented. LLMs cannot determine what fields to expect from tool responses (e.g., what does 'ls' return? Are they objects with name/size/date, or strings?). All 17 tools lack output schema documentation.
Destructive tools (rm, touch overwrite, mv) lack confirmation or dry-run patterns. 'rm' with a path parameter could delete critical files without reversibility guards. No confirm_before_execute mechanism.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 65 | <=2025-11-25 | v2 |
| 2026-03-09 | D | 59 | - | v1 |
Retrieves the content of the currently active page. Extracts content from the page using a CSS selector. Can return either plain text (with readable formatting) or raw HTML source code.
List directory contents with details (size, time). Recommended for checking specific folders. For overview, use 'tree' instead.
Create a new directory (recursive).
Move or rename a file or directory.
Opens a specific URL in a new browser tab. This tool navigates to the specified URL and optionally waits for network requests to complete before returning. Supports both HTTP and HTTPS protocols.
Remove a file or directory. [WARNING] This action is IRREVERSIBLE. Check contents with 'ls' or 'tree' before deleting directories. Do not delete files unless explicitly requested by the user.
Search for a text pattern in files within a directory used for grep. [Hint] Use specific, unique keywords. Common words will return too many results.
Captures a full screenshot of the current system desktop and saves it to a local path.
Create or update a file with content. [CAUTION] This will OVERWRITE existing files completely. To add content, use 'append' instead.
Recursively list directory contents in a tree structure. [IMPORTANT] - ALWAYS specify a constrained 'max_depth' (e.g. 2 or 3) to prevent context overflow. Default is depth 2. Use -1 only if you are sure the directory is small.
Many parameters lack descriptions entirely or are under-described. File operations (ls path, cat path, tree path) do not explain whether paths are relative/absolute, whether they support globs, or what errors occur if the path doesn't exist. mkdir/rm path descriptions are minimal.
No error handling guidance. Tools do not document what to do if a file is not found, if permissions are denied, or if an operation times out. No recovery hints for LLMs.
Parameter constraints not enforced or documented. 'max_length' in get_clipboard (default 2000) is mentioned, but range/min/max for numeric params is absent. 'max_depth' in tree defaults to 2 but has no stated maximum. 'timeout' values in open_url and get_page_content are unconstrained.
Browser-based tools (open_url, get_page_content) do not document timeout behavior or what happens if a page is blocked. Selectors in get_page_content may not match, no guidance on fallback or error recovery.
take_screenshot and get_page_content do not indicate whether they interact with a GUI running on the local machine. This is a critical security/capability gap, agents may assume they can screenshot any system, but implementation likely only works if a display server is available.