Loading ToolBench
MCP server for AI-powered consultation. GPT-5 strategic planning, Perplexity real-time search, GPT Image visual creation, with intelligent context gathering
Kortx MCP has well-structured tool definitions with clear names and solid descriptions, but suffers from critical gaps in schema completeness and parameter documentation. All 8 tools follow verb_noun naming conventions and have 100-200 character descriptions (baseline compliant). However, the source code provided does NOT show complete input schemas with all parameter type definitions, only high-level descriptions of parameters are visible. For example, 'think-about-plan' shows a schema object with properties like 'plan', 'context', 'preferredModel', but many tools lack visible minLength/maxLength constraints, and the 'create-visual' tool parameters ('size', 'quality', 'background', 'outputFormat') have NO type declarations at all in the visible schema. Additionally, NO tool shows an output schema definition, which violates the pattern:tool and pattern:tool-description baselines requiring documented return types. Tools are READ_ONLY (good for safety), but risk classification is not clearly communicated in descriptions. The codebase shows sophisticated LLM integration and caching (BaseTool class), suggesting production maturity, but the visible tool metadata is incomplete.
Execute multiple consultation tools in parallel for comprehensive analysis. Combines insights from think-about-plan, suggest-alternative, improve-copy, solve-problem, consult, search-content, and create-visual.
General-purpose consultation tool for asking GPT-5 strategic questions. Gathers context automatically and provides thoughtful analysis.
Generate images using GPT Image model. Creates custom visuals from text descriptions with extensive customization options.
Improve written content by enhancing clarity, engagement, tone, and effectiveness. Analyzes structure, grammar, and persuasiveness.
Search the web in real-time using Perplexity AI. Returns relevant information with sources, citations, and optional images.
Output schemas NOT documented for any tool. LLMs cannot infer what fields to expect from responses, preventing downstream tool chaining and forcing agents to parse unstructured results.
'create-visual' parameters (size, quality, background, outputFormat) lack type definitions. Schema shows only parameter names with descriptions but no 'type' field, violating JSON Schema basics and making the tool's input contract ambiguous.
'batch-consult' tool accepts a 'tools' array parameter but does NOT document which tool names are valid, what happens if invalid tool names are passed, or whether all tools must exist. This invites hallucinated values.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 58 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 23 | - | v1 |
Get structured problem-solving guidance. Breaks down complex problems, identifies root causes, evaluates solutions, and provides implementation steps.
Suggest alternative approaches or solutions. Considers different paradigms, simpler solutions, proven patterns, and trade-offs.
Get strategic feedback on plans and approaches. Analyzes clarity, feasibility, risks, dependencies, and suggests alternatives.
No output schema examples or field documentation. While descriptions explain WHAT tools do, LLMs lack clarity on WHAT FIELDS to extract from responses. A response from 'search-content' may include 'sources', 'citations', 'images', but the tool definition never confirms this.
Tool descriptions do not explicitly state that all tools are READ_ONLY (safe to call repeatedly). While Risk field shows 'READ_ONLY', this metadata is not visible in the LLM-facing description text. Agents need this guidance in the description itself.
'preferredModel' enum on multiple tools references GPT-5 variants (gpt-5, gpt-5-mini, gpt-5-nano, gpt-5.1-2025-11-13, gpt-5.1-codex) that do not exist in OpenAI's public API as of 2025. These are either aspirational, internal, or hallucinated. If real, document; if not, replace with actual available models (gpt-4o, gpt-4-turbo).
Parameter 'minLength' constraints are present (e.g., query minLength=5 in 'consult', plan minLength=10 in 'think-about-plan'), but NO maxLength limits are defined. Unbounded string inputs can cause token explosion and context window exhaustion.
No error handling guidance documented. When an LLM calls these tools and gets an error (e.g., Perplexity API rate limit on 'search-content', OpenAI quota exceeded on 'create-visual'), there is no documented recovery path or retry guidance in the tool description.