MCP Hive shows ambitious swarm orchestration architecture but suffers from critical naming, schema, and composition gaps that severely limit LLM usability. Tool names use noun-first patterns (hive_swarm, hive_harvest) rather than action-verb conventions, violating the 90% A+ baseline of verb-starting names. Parameter descriptions exist but are often vague (e.g., 'The content of the insight' for a complex object). Error handling returns synthesized batches on failure but lacks recovery guidance or actionable next steps. Most critically, three of four tools (hive_post_insight, hive_spawn_subtask, and partially hive_swarm) rely on runtime context (swarm_id, task_id) that is not reliably produced by other tools in the suite, breaking tool-chain composability. The server attempts sophisticated agent-to-agent delegation but exposes this complexity to the LLM caller rather than abstracting it.

• Rename tools to verb-first patterns: hive_swarm → orchestrate_swarm or spawn_swarm; hive_harvest → collect_swarm_results or get_swarm_results; hive_post_insight → share_discovery or report_finding; hive_spawn_subtask → create_subtask or delegate_specialist.
• Document output schemas explicitly in tool descriptions and in code. Example: 'Returns { swarm_id: string, total_agents: number, successful: number, results: Array<{role: string, status: 'success'|'fatal_error', findings: Array<{type, description, severity}>} }'.
• Replace unconstrained content and context parameters with typed sub-objects. Example: content should have explicit fields like { discovery_title: string, finding_description: string, evidence: string } instead of z.record(z.string(), z.any()).
• Add actionable error messages. Instead of 'Orchestration aborted', return 'Orchestration failed: {reason}. Retry with valid paths or simplify tasks. Available help: (error details)'.
• Document interdependencies explicitly. Add to hive_spawn_subtask description: 'Use this only within an active swarm started by orchestrate_swarm(). Requires the swarm_id returned by that tool.'
• Validate and sanitize path inputs. Check for traversal sequences (..), symlinks, and existence before using. Return 'Invalid path: {reason}' with guidance.
• Add pagination and result-limiting for harvest. Document: 'Returns up to 50 most recent results. Use offset/limit for large swarms.' Include total_results and next_cursor in response.
• Specify task_type enum in hive_spawn_subtask instead of free-form string. Example: task_type enum: ['sql_expert', 'security_auditor', 'performance_analyst', ...] or document format rules.