Loading ToolBench
MCP server for sandboxed code execution with context window reduction
mcp-exec has 4 tools with mixed quality. The 'exec' tool has a well-structured schema with runtime options and timeout support, but descriptions are generic and lack LLM-optimized guidance on when/why to use each tool. The 'tools' search utility has a clear purpose but minimal parameter documentation. 'echo' and 'add' are test fixtures with trivial descriptions. No tool descriptions explain error recovery, prerequisites, or dependencies. Parameter descriptions are sparse, 'query' in 'tools' lacks format guidance; 'runtime' in 'exec' lacks explanation of when to choose node vs bash vs python. Output schemas are not documented anywhere. Error handling is absent from descriptions. Overall: functional but below production baseline (194 char avg description; this server averages ~60 chars).
Adds two numbers and returns the sum
Returns the input message unchanged
Execute code in a sandboxed environment. Only the return value (Node) or stdout (Bash/Python) enters the context window — intermediate data stays in the sandbox.
Search available MCP tools. Pass "*" to list all tools, or a query to filter. Returns trimmed summaries — full schemas are never loaded into context. Unavailable servers appear with status "unavailable" and a reason.
Tool descriptions are too short and lack LLM-optimized guidance. 'tools' (67 chars) and 'exec' (102 chars) fall below the 194-char production baseline. Descriptions do not explain WHEN to use each tool, prerequisites, or error recovery paths.
Parameter descriptions are missing or minimal. 'query' in 'tools' lacks format guidance (e.g., 'wildcard * lists all; partial strings filter by name'). 'runtime' in 'exec' does not explain when to choose node vs bash vs python or what each runtime is optimized for.
Output schemas are not documented. LLMs cannot plan downstream calls or extract required fields. 'exec' returns stdout/return values but the response structure is not specified. 'tools' returns 'trimmed summaries' but the schema is invisible.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
Error handling is absent from tool descriptions. No guidance on retryability, user-fixable errors, or recovery steps. 'exec' can timeout or fail but descriptions do not explain what to do next.
Test fixtures 'echo' and 'add' have trivial descriptions (20 chars each). These should either be removed from production or upgraded with proper documentation explaining their role in the server.