MCP server for interacting with Defense.com API to retrieve threat intelligence, manage detections, and provide security training modules
Defense MCP has moderate definition quality with consistent issues across naming, descriptions, and schema completeness. All 8 tools have basic names starting with action verbs (get_, create_) and descriptions present, but descriptions are uniformly brief (50-90 chars) and lack depth about WHEN to use tools or prerequisites. Input schemas exist for all tools and include parameter descriptions, but are minimal, no enum constraints, no format specifications, and no output schema documentation. Parameters like 'threat_id' and 'session_id' lack type suffixes despite being identifiers. The 'create_training_session' tool has ambiguous interdependent parameters (threat_id vs threat_category) with insufficient documentation of mutual exclusivity. No error handling guidance visible. The server provides read-only operations and training session management, but lacks the depth of documentation needed for confident LLM tool selection in production scenarios.
Create a new security training session based on threat context or threat category
Get the list of available training categories
Fetch detections from the Defense API with optional filtering
Get detailed information about a specific threat by ID
Fetch threats from the Defense API with optional filtering and pagination
Get the current question in a training session
Get the progress of a training session including current question, score, and completion status
Descriptions are uniformly brief (35-50 chars) and lack context for tool selection. 'Fetch threats from the Defense API with optional filtering and pagination' does not explain WHEN to use this vs get_threat_details, WHAT filtering is available, or typical use cases.
No output schemas documented. Tools return structured data (pagination info, threat objects, training session details) but LLMs cannot infer what fields to expect. Missing fields like pagination cursors, threat severity levels, or training score formats force agents to guess.
create_training_session has ambiguous parameter interdependencies: threat_id and threat_category are mutually exclusive but descriptions do not state this. LLMs will attempt to pass both, causing tool failure. Parameter descriptions lack guidance on which to prefer or when each applies.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 40 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 25 | - | v1 |
Submit a response to a training question and get feedback
No enum constraints on constrained parameters. 'threat_category' accepts 'ransomware, malware, phishing, vulnerability_management, network_security, general' but is defined as a free-form string. LLMs will hallucinate invalid values like 'data-theft' or 'ransomware_attack'. Should use JSON Schema enum.
No error handling guidance visible in tool definitions. No indication of what happens on invalid threat_id, missing training session, or API failures. No recovery patterns like 'If threat not found, try search_threats()' documented.
Parameters use bare identifier names without type suffixes. 'threat_id', 'session_id', 'user_id' lack clarity on format (UUID, integer, string slug). Ambiguous formats invite LLM errors.
'days_back' parameter in get_threats and get_detections lacks numeric constraints (min/max). LLMs could pass 999 or 0, breaking the tool. Should specify '1-365 days' or similar.
'get_all_pages' parameter defaults to false but lacks guidance on pagination behavior. Does it return only page 1 by default? How many results? Does it support per_page limits? Undocumented pagination forces agents to guess.
No tool composition guidance. How do get_threats and get_threat_details relate? Should an agent call one then the other, or is get_threats sufficient for most tasks? Unclear tool boundaries force suboptimal agent planning.