Loading ToolBench
Multi-provider LLM Delegator MCP Server that delegates tasks to specialized expert personas (architect, code_reviewer, security_analyst, plan_reviewer, scope_analyst) with caching, memory, and background job support
Server has 11 tools with schemas and descriptions present, but quality is inconsistent. Tool names follow verb_noun convention (good), but descriptions are generic and lack LLM-optimization guidance. Most parameters have types and descriptions, but lack constraints (enums, ranges, patterns). Output schemas are not documented. Error handling is minimal, responses lack recovery guidance. Expert delegation tools (glm_architect, glm_code_reviewer, etc.) are nearly identical, violating single-responsibility principle. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Baseline: 549 production tools average 194-char descriptions; these average ~120 chars and lack specificity on WHEN to use each tool vs alternatives.
Delegate to the Architect expert (glm-5)
Get caching and memory system statistics
Delegate to the Code Reviewer expert (glm-5)
Use LLM to enhance a prompt before sending to expert. Improves clarity, structure, and completeness.
Retrieve the result of a background job. Blocks until job completes (long-polling).
Delegate to the Plan Reviewer expert (glm-5)
Intelligently route a task to the best available tool (expert/skill/team) using hybrid keyword + semantic matching
Five expert delegation tools (glm_architect, glm_code_reviewer, glm_security_analyst, glm_plan_reviewer, glm_scope_analyst) are identical in schema and differ only in description. This violates single-responsibility and forces LLMs to reason about subtle distinctions. Descriptions do not explain WHEN to choose one expert over another or what each specializes in.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels declared in source (READ_ONLY, WRITE). glm_workflow is marked WRITE but has no destructiveHint. Agents cannot distinguish safe from risky operations.
Descriptions lack LLM-optimization guidance. Generic phrases like 'Delegate to the Architect expert' do not explain WHEN to call this tool, what it returns, or how it differs from alternatives. Baseline: A+ tools average 194 chars with explicit WHAT/WHEN/RETURN structure.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | <=2025-11-25 | v2 |
Delegate to the Scope Analyst expert (glm-5)
Delegate to the Security Analyst expert (glm-5)
Validate a prompt using static rules (no LLM call). Checks file existence, hallucination signals.
Execute automated workflow: Code → Review → Test → Fix → Report with state machine management
Output schemas are not documented. Tool handlers return JSON with fields like 'content', 'isError', 'message', but LLMs have no schema to parse responses. Agents must infer structure from examples, increasing hallucination risk.
Error handling lacks recovery guidance. Handlers catch exceptions and return generic error JSON, but do not tell LLMs what to do next (retry, ask user, abort). Baseline: A+ tools include 'Try search_users() with a partial name' style guidance.