Loading ToolBench
A local MCP (Model Context Protocol) server that acts as a translation layer between MCP requests and various LLM provider APIs (OpenAI, Anthropic). Supports chat completions, embeddings, and tool execution with streaming capabilities.
Server provides 12 tools with basic schemas and descriptions, but quality is inconsistent. Most tools have descriptions (good), but many lack parameter-level detail and output schema documentation. Tool naming follows verb_noun convention (positive). However, parameter descriptions are sparse or generic, and error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint). Schemas present but incomplete, many parameters lack type constraints, enums, or validation rules. Output structures undocumented. This is typical of community-built servers: functional but not production-grade.
Perform chat completion with support for multiple LLM providers (OpenAI, Anthropic) with optional tool calling
Compute importance scores for nodes using graph algorithms (pagerank, betweenness, closeness, degree)
Create embeddings for input text using specified embedding model
Execute a specified tool with given parameters
Find communities most relevant to the query from the knowledge graph
Find entities most similar to the query embedding from the knowledge graph
Parameter descriptions are generic or missing detail. E.g., 'tools' param in chat_completion says 'Optional tools/functions available to the model' but does not specify schema, format, or how the model should interpret them. LLMs cannot infer structure from vague descriptions.
Output schemas are not documented. Tools like 'find_relevant_entities', 'generate_graph_context', and 'hybrid_search' do not specify what fields they return, their types, or structure. LLMs cannot plan downstream calls or extract data without knowing the response shape.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). The 'execute_tool' is marked WRITE but has no annotation in the schema. Streaming tools lack idempotent hints. Current MCP spec (2026-07-28) expects these for agent safety.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | <=2025-11-25 | v2 |
Generate comprehensive context for Graph RAG including entities, communities, relationships, and neighborhood
Get the neighborhood subgraph around specified entities
Get list of supported models for the configured LLM providers
Check server health and provider availability status
Perform hybrid search combining vector similarity and graph structure for improved relevance
Stream chat completion responses from LLM providers with chunked output
Error handling lacks recovery guidance. No tool description explains what to do if a call fails (e.g., 'If model not found, call get_supported_models() first'). Error responses are not shown in the schema.
Graph RAG tools (find_relevant_entities, find_relevant_communities, etc.) accept 'query_embedding' as a required array parameter but do not specify: vector dimension, format (float32 vs float64), or how to compute it. LLMs cannot call these tools without external embedding logic.