MCP server for Apstra AOS (Automated Operational System) integration, providing tools to manage blueprints, racks, routing zones, virtual networks, and deployment operations
This server has foundational issues across naming, descriptions, and schema completeness. While tool names follow the verb_noun pattern (get_, create_, delete_, list_), most descriptions are extremely brief (under 50 chars) and lack context about when to use each tool or what they return. Parameter descriptions exist for most tools but are minimal. Output schemas are not documented anywhere. Error handling is present but returns unstructured responses without guidance on retry logic or recovery. The server exposes credentials as global constants rather than using secret injection, creating a security vulnerability. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present despite clear risk classifications. Schemas are properly typed (string, integer) but lack enums, ranges, or constraints on parameters like 'template_name' or 'staging_version'.
Creates a new blueprint using a specified template.
Creates a virtual network in a given blueprint and routing zone
Delete the blueprint from Apstra Instance
Deploys the config for a blueprint
Gets blueprint information
Gets the diff status for a blueprint
Gets rack information for a blueprint
Credentials stored as global constants (aos_server, username, password) instead of server-side secret injection. Credentials appear in source code and could be logged in agent traces.
Tool descriptions are extremely brief (20-45 chars), lacking context about when to use each tool, what it returns, and when it modifies state. get_bp ('Gets blueprint information'), deploy ('Deploys the config for a blueprint'), and delete_bp ('Delete the blueprint from Apstra Instance') all lack clarity on prerequisites or output structure.
No output schemas documented. LLMs cannot predict what fields to expect from responses (e.g., does get_bp return a list of blueprint objects? Does each have an 'id' field? Does get_racks return rack count?). This forces agents to guess and risks downstream tool failures.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 40 | 2026-07-28+ | v2 |
Gets routing zone information for a blueprint
List all virtual networks in a blueprint
No tool annotations for risk levels. Destructive tools (delete_bp, deploy) and write tools (create_vn, create_blueprint_from_template) lack readOnlyHint, destructiveHint, or idempotentHint annotations. Agents cannot infer safety constraints or retry behavior.
Error handling returns unstructured dicts or None without recovery guidance. E.g., auth() returns None on failure, but tools don't explain what went wrong or what the agent should try next. delete_bp returns {'error': '...', 'status': 'failed'} but lacks retry classification.
Parameter 'template_name' in create_blueprint_from_template has a default value 'USE_YOUR_OWN_TEMPLATE' and no enum of valid template names. LLMs cannot discover valid templates without calling list_templates (which doesn't exist) and will likely pass the default or hallucinated values.
Parameter 'staging_version' in deploy is an integer with no description of valid range. Can an agent pass any integer? Is there a minimum/maximum? This invites out-of-bounds values.
Tool names 'get_rz' (routing zones, abbreviated) is ambiguous. RFC-style abbreviations are harder for LLMs to parse than full names. Use 'get_routing_zones' or 'get_security_zones' depending on the API semantics.
No confirmation or dry-run support for destructive operations. delete_bp immediately deletes without a separate confirm step. Agents making mistakes will destroy blueprints irreversibly.
API responses are not filtered. Tools return full response.json() without stripping internal metadata, IDs the user never needs, or audit fields. This wastes tokens and dilutes signal. Example: create_vn likely returns entire VN object with internal fields.