Loading ToolBench
Autonomous swarm intelligence engine exposing multi-agent simulation and forecasting capabilities via Model Context Protocol. Allows MCP-compatible hosts to run simulations, query past runs, and retrieve reports.
HIVE MCP server has 6 tools with basic schemas and descriptions, but significant gaps in parameter documentation, output schema clarity, and error handling guidance. Tool names follow verb_noun convention (run_simulation, list_runs, get_run, get_report, search_web, search_news), which is good. However, descriptions are generic (avg ~80 chars, below 194-char baseline), parameters lack type info and descriptions in several cases, and output schemas are not documented. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all tools being READ_ONLY. Error handling is minimal, no recovery guidance or actionable error messages visible in code. The tool registry (BaseTool, PythonFunctionTool, HttpApiTool, McpClientTool) is well-structured internally, but the MCP server integration (hive_mcp_server.py) is truncated in the provided source, preventing full assessment of tool registration and response formatting.
Get the full markdown report for a specific simulation run.
Get full details for a specific simulation run including verdict, agents, debate, and scenarios.
List all previous simulation runs with their goal snippets and probability scores.
Run a full autonomous swarm forecast on any goal. Returns a comprehensive verdict with probability scores, predicted developments, risk analysis, and narrative forecast.
Search recent news articles.
Search the live web for current information.
Output schemas not documented. Tools return results but LLMs cannot predict field structure (e.g., what fields does run_simulation verdict contain? Is it a string, object, or array?). This forces agents to guess and wastes tokens on parsing.
Parameter descriptions missing or minimal. 'goal' in run_simulation has a description, but 'constraints' and 'timeline' are vague ('Optional constraints (solo, budget, runway, etc.)', what format? comma-separated? JSON?). 'query' in search_web/search_news has no description at all.
Tool descriptions are generic and under 100 chars (baseline 194). 'List all previous simulation runs with their goal snippets and probability scores' (list_runs) does not explain WHEN to call it vs get_run, or what 'goal snippets' means. Agents cannot distinguish between similar tools.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
No tool annotations despite all tools being READ_ONLY. MCP spec supports readOnlyHint, destructiveHint, and idempotentHint. Agents cannot infer safety properties without explicit annotations.
Error handling in tool code (HttpApiTool, McpClientTool) returns raw HTTP status codes and truncated error text. No recovery guidance (e.g., 'HTTP 404: not found' vs 'Run_ID not found. Call list_runs() to see available IDs'). Agents cannot self-correct.