Governed Microsoft Graph agent and MCP server for Microsoft 365, Power Platform, Intune, and Windows integration with fail-closed tool policies
This server has serious definition quality gaps across multiple dimensions. While some tools have descriptions, many lack sufficient schema clarity, parameter descriptions are sparse or missing entirely, and error handling guidance is absent. The server mixes read-only and write operations without clear idempotency markers. Of 15 tools, only ~3 have reasonably complete definitions (get_document_capabilities, generate_word_document, generate_powerpoint_presentation). The remaining 12 tools suffer from incomplete parameter documentation, missing schema validation details, and no guidance on output structure. Complex tools like microsoft_applications and microsoft_domains accept a generic 'params_json' string parameter, which is a major anti-pattern, this forces LLMs to construct arbitrary JSON without schema validation. The server exposes internal Microsoft Graph complexity without abstracting it into agent-friendly interfaces.
Create a OneNote page from the raw HTML body required by Graph.
Generate a non-macro PowerPoint .pptx presentation from validated slides, metadata, or a template confined to the configured root.
Generate a non-macro Word .docx file from validated paragraphs, tables, metadata, or a template confined to the configured root.
Report whether local Word and PowerPoint OOXML generation backends are installed. This tool does not require Microsoft authentication.
Get Excel table.
Get Excel workbook.
Generic params_json pattern breaks schema validation (microsoft_applications, microsoft_domains). LLMs must construct arbitrary JSON without type checking, enums, or constraints. This is a critical anti-pattern.
Excel and OneNote tools (8 tools) have trivial descriptions (10-20 chars: 'Get Excel workbook.', 'List Excel worksheets.'). LLMs cannot determine when to use these vs alternatives or what to expect. Baseline is 194 chars average for A-tier tools.
Parameters lack descriptions across Excel/OneNote tools. 'drive_id', 'item_id', 'worksheet_id' are passed with only type info ('Drive identifier'), but LLMs need context: where do these IDs come from? How do I obtain them? This forces extra discovery calls.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | <=2025-11-25 | v2 |
Get Excel worksheet.
Get Onenote page content.
List Excel tables.
List Excel worksheets.
Return only keyed opaque nodes and structural relationships. Raw Microsoft identifiers and content never appear in the returned source connector projection. The deployment-owned pseudonymization key is required.
List Onenote notebook sections.
List Onenote section pages.
Manage microsoft applications operations.
Manage microsoft domains operations.
No output schemas documented. Callers cannot see what fields to expect from list_excel_tables, get_onenote_page_content, etc. LLMs must infer structure, risking parsing errors and wasted context.
No pagination documented for list_* tools. list_excel_tables, list_excel_worksheets, list_onenote_section_pages lack offset/limit/next_cursor parameters and total count in response. Large result sets will blow context windows.
microsoft_applications and microsoft_domains accept string action parameters with no enum constraint. Actions like 'list_applications', 'create_domain', etc. are only documented in the description text. LLMs cannot autocomplete or validate without parsing text.
No error handling guidance. Tools offer no recovery hints (e.g. 'If item not found, call search_excel_files()' or 'Invalid worksheet_id, try list_excel_worksheets first'). Agents stuck on failures.
Document generation tools (generate_word_document, generate_powerpoint_presentation) lack examples of valid request structures. WordDocumentRequest and PowerPointPresentationRequest types are referenced but not expanded in the schema. LLMs cannot construct valid payloads.
list_microsoft_ingestion_projection description mentions pseudonymization and OData but does not explain what 'kind: messages' vs 'files' returns, what fields are included, or when to call this vs other tools. LLMs guessing at intent.
No tool annotations visible (readOnlyHint, destructiveHint, idempotentHint). write operations like generate_word_document, create_onenote_page lack destructiveHint=true, making it unclear to agents which tools have side effects and which are safe to retry.