Loading ToolBench
Asynchronous job dispatch and management for Claude Desktop, with detached worker processes that survive server restarts. Supports both STDIO and HTTP/HTTPS transports.
Three tools with basic schemas and descriptions, but significant gaps in parameter documentation and error handling. Tool names follow verb_noun convention (claude_start, claude_check, claude_jobs), which is good. However, descriptions are minimal (50-80 chars), parameter descriptions are sparse or missing, and no output schemas are documented. The 'effort' parameter in claude_start lacks an enum constraint despite having a fixed set of values (low, medium, high, xhigh, max). No error recovery guidance. No tool annotations (readOnlyHint/destructiveHint). This is typical community-server quality, functional but not production-grade.
Check the status of a dispatched job by jobId. Returns current state, exit code, and log output.
List all dispatched jobs with their current states and metadata.
Dispatch an asynchronous job to run a command with the Claude CLI. Returns a jobId for tracking.
Effort parameter lacks enum constraint. Accepts 'low|medium|high|xhigh|max' but defined as free-form string, inviting hallucinated values.
No output schemas documented. LLMs cannot predict what claude_check or claude_jobs return (fields, types, structure), forcing guesswork on downstream tool calls.
Descriptions are too brief (30-80 chars). Lack context on WHEN to use each tool and what distinguishes them. 'List all dispatched jobs' does not explain when to call this vs claude_check.
No error handling guidance. If a jobId is invalid or a job fails, no recovery path is documented. LLM receives raw error with no actionable next step.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
No tool annotations. claude_start is destructive (WRITE risk) and should carry destructiveHint=true. claude_check and claude_jobs are read-only and should carry readOnlyHint=true.