Model Context Protocol server for Autodesk Construction Cloud Issues - enables natural language interaction with ACC Issues API through Claude AI
This MCP server has significant quality gaps across naming, descriptions, and schemas. Of 6 tools, most have acceptable names but lack depth in descriptions. Input schemas are present with type information but parameter descriptions are generic and insufficient. Output schemas are completely undocumented, critical for LLM reasoning. Error handling guidance is absent. The server exposes operation semantics (WRITE vs READ_ONLY) but provides no recovery paths or actionable error messages. No tool annotations (readOnlyHint/destructiveHint) are present despite tools having clear risk profiles.
Add a comment to an issue
Create a new issue
Generate a report of issues
Get a specific issue by ID
Search for issues matching specified filters
Update an existing issue
No output schemas documented. LLMs cannot predict response structure, forcing them to parse unstructured output and blocking downstream tool chaining. For example, create_issue should document what fields are returned (issue_id, created_timestamp, status, etc.) so add_comment can reference the new issue_id.
Descriptions are too brief and generic. 'Get a specific issue by ID' (27 chars) and 'Search for issues matching specified filters' (45 chars) lack actionable context. Example: get_issue_details should mention whether it returns linked issues, attachments, or comment counts; search_issues should explain pagination behavior and default result limit.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 48 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 42 | - | v1 |
Parameter descriptions lack detail. 'region' appears across all tools but is described as 'Optional region parameter' with no explanation of valid values, format, or default. 'status' in search_issues is an array but no enum constraint provided, LLM will hallucinate invalid statuses. 'report_type' in generate_report has no enum or valid examples.
No tool annotations despite clear risk profiles. Tools tagged WRITE (create_issue, update_issue, add_comment) lack destructiveHint annotation; READ_ONLY tools lack readOnlyHint. This prevents agents from reasoning about retry safety and side effects. Per spec 2026-07-28, tool annotations enable proper agent planning.
No error handling guidance. Tools declare no possible error conditions or recovery steps. If search_issues returns empty, should the agent try broader filters? If create_issue fails, is it retryable? Without explicit error classification (retryable, user-fixable, fatal), agents cannot self-correct.
Parameter naming inconsistencies and missing type hints in some contexts. create_issue accepts 'assigned_to_type' but update_issue does not, why the difference? What are valid values for assigned_to_type? Date parameters (due_date, start_date, created_after, etc.) lack format specification (ISO 8601 expected but not stated).
Pagination not fully specified. search_issues accepts 'limit' and 'page' parameters but no documentation of max limit, default page size, or total count in response. Per pattern:paginated-result, tools returning lists must support pagination and return total count or next_cursor.
Tool chaining blocked by missing response field documentation. If create_issue succeeds, does it return an issue_id? If add_comment is called immediately after, how does the agent reference the new issue? Without documented output fields, agents cannot chain operations.