Loading ToolBench
Advanced cognitive thinking MCP with DAG-based thought graph, 10 reasoning strategies, node aliases, session persistence, structured responses, auto strategy selection, and 17 tools with help & conclude
Server has 18 tools with mostly complete schemas and descriptions. Naming conventions are verb-based and reasonably clear (think, graph, critique, prune, etc.). Descriptions are substantive (100-300+ chars typical) and explain WHEN to use each tool. However, there are critical gaps: (1) Many parameter descriptions lack detail about format constraints, valid ranges, or dependencies. (2) Output schemas are not documented, descriptions mention what the tool returns, but formal response schemas are absent from the definition. (3) Error handling is not explicit in tool definitions, no guidance on what to do if a call fails. (4) Tool compositions have hidden dependencies (e.g., 'think' with strategy='auto' triggers internal logic not explained to the LLM). (5) Some tools like 'ethics' have minimal descriptions relative to their scope. The server would work for basic use but lacks production-grade precision.
Generate cross-domain analogies and apply them to the current reasoning. Maps problem structures across disciplines for novel insights.
Generate a conclusion from the current thought graph. Analyzes the best path through the graph, evaluates confidence, and produces a structured summary with supporting evidence.
Generate a critical analysis of a specific thought node or the entire graph. Evaluates logical consistency, confidence, supporting evidence, and suggests improvements.
Generate counterarguments and alternative perspectives. Applies dialectical opposition to strengthen reasoning through critical challenge.
Create explicit edges between nodes to build relationships. Supports derives_from, contradicts, supports, refines, challenges, synthesizes, parallels, abstracts, and instantiates edge types.
Analyze emotional and interpersonal dimensions of reasoning. Evaluates emotional coherence, social impact, and emotional intelligence aspects.
Output schemas not documented. Tools describe what they return in text (e.g., 'visualize the thought graph') but do not provide formal JSON Schema definitions for response structures. LLMs cannot reliably parse unstructured output or plan downstream calls without knowing expected fields.
Parameter constraints not fully specified. Many parameters lack descriptions of valid ranges, formats, and dependencies. For example, 'depth' in critique is documented as 1-5 but no guidance on what depth values mean semantically. 'confidence' in think accepts 0-1 but no explanation of how it influences reasoning.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2025-06-18+ | v2 |
Evaluate reasoning for ethical implications, biases, and stakeholder impacts. Provides structured ethical analysis and recommendations.
Retrieve and visualize the current thought graph structure with nodes, edges, branches, and statistics. Supports multiple output formats including ASCII and SVG visualization.
Display help information about available tools, strategies, thought types, and how to use the deep-thinker MCP.
Integrate external knowledge, facts, and constraints into thought nodes. Supports validation, gap detection, and consistency checking.
Analyze and remove low-confidence, redundant, or dead-end nodes from the thought graph. Returns analysis of prunable nodes and optimization results.
Reset the cognitive state and clear the thought graph. Creates a new thinking session with optional problem statement preservation.
Search the thought graph by type, tag, content, or path. Returns matching nodes and their relationships.
Get the current cognitive state including strategy, progress metrics, stagnation detection, and metacognitive analysis.
Add or manage tags on thought nodes for categorization and filtering.
Project thoughts temporally into past or future. Analyzes how reasoning evolves over time, predicts outcomes, and identifies temporal dependencies.
Add a thought to the cognitive graph using the current strategy. Supports sequential, dialectic, parallel, analogical, and abductive reasoning strategies. Each thought becomes a node in a DAG with confidence scoring, edges, and metadata.
Visualize the thought graph in different formats (ASCII, SVG) with customization options.
Error handling not documented. Tool definitions include no guidance on error conditions, recovery steps, or what the LLM should do if a call fails. For example, what happens if 'parentId' references a non-existent node? Should the LLM retry, fallback, or ask the user?
Hidden parameter dependencies and side effects. The 'think' tool's strategy='auto' triggers metacognitive logic ('automatic strategy selection') not explained in the description. The interaction between 'parentId', 'branch', and strategy is undocumented. LLMs cannot reason about these couplings without explicit guidance.
Minimal descriptions for specialized tools. 'ethics' tool description is only ~100 chars and does not explain when to use it, what frameworks are available beyond the enum, or what 'balanced' means as a default. 'emotional_intelligence' similarly lacks context.
Nested objects lack full schema definition. Parameters like 'dialectic', 'parallel', 'analogical', 'abductive', 'firstPrinciples', 'counterfactual', 'systemsThinking' define nested properties but no guidance on mutual exclusivity, required fields within each, or validation rules. LLMs may construct invalid nested objects.
State mutation not clearly marked. Tools like 'prune', 'knowledge', 'tag', 'edge', 'reset' modify graph state, but descriptions do not explicitly state this is destructive or idempotent. LLMs need to know whether retrying is safe.
Tool composition unclear. Multiple tools operate on the same graph (think, edge, tag, knowledge, prune) but descriptions do not explain interaction order or dependencies. For example, should 'tag' be called after 'think', or can they be interleaved?