The server implements 24 tools with basic schemas and descriptions, but critical gaps in parameter validation, error handling, and output documentation severely limit production readiness. Tools follow a clear naming convention (pbir_*, manage_*, list_*, etc.) but descriptions are often too generic (e.g., 'Debugs the folder structure' for pbir_inspect_structure), and many lack actionable guidance on when to use them. Parameter schemas exist but lack depth: most omit type constraints (enums for operation choices), valid ranges (e.g., pixel bounds for layout), or clear format specifications. Output schemas are entirely undocumented, no tool declares what fields or structure it returns, forcing LLMs to guess at chaining parameters. Error handling is absent; no tool shows how to recover from common failures (connection lost, invalid field reference, permission denied). The server uses STDIO transport, which caps it at 50 regardless of other quality factors.
Tools (24)
get_model_inforead onlysource verified62/100
Get overall model metadata (tables, measures count, etc).
get_vertipaq_statsread onlysource verified62/100
Get VertiPaq/compression statistics for the model.
list_objectsread onlysource verified62/100
List objects: tables, measures, columns, relationships, roles, partitions.
Output schemas completely undocumented. No tool declares return type or field structure. Forces LLMs to guess at chaining parameters and invites context loss between calls.
No error recovery guidance. Tools like pbir_create_page, run_dax, and manage_table lack error descriptions or suggestions for recovery (e.g., 'If connection fails, check that Power BI Desktop is running'). LLMs receive bare error codes with no actionable next step.
Recommendations
Document output schema for every tool. Include example response JSON showing field names, types, and structure. For list tools, show total count and next_cursor if pagination is needed. Example for pbir_get_info: '{"status": "detected", "project_path": "/path/to/project", "pages": [{"name": "Page1", "visuals_count": 5}, ...], "error": null}'
Add enum type constraints to all 'operation' parameters. Example: manage_measure with operation: {type: 'string', enum: ['create', 'update', 'delete'], description: 'Operation to perform: create=new measure, update=modify existing, delete=remove measure'}. This prevents hallucinated operation names.
Replace 'Properties formatting as JSON' (pbir_format_visual) with a typed object listing valid properties and their types. Example: properties: {type: 'object', properties: {fontSize: {type: 'number', description: 'Font size in points, 8-72'}, color: {type: 'string', pattern: '^#[0-9A-F]{6}$', description: 'Hex color code'}}}
Extend parameter descriptions with format and constraint rules. Example for pbir_create_page: name: {type: 'string', description: 'Name of the page (1-100 characters, unique within project, no special chars except underscore)'} and pbir_update_visual_layout: x: {type: 'integer', minimum: 0, maximum: 9999, description: 'X position in pixels'}, width: {type: 'integer', minimum: 50, maximum: 9999, description: 'Width in pixels, minimum 50'}
Add error recovery guidance to every tool description. Examples: 'If connection_error is returned, ensure Power BI Desktop is running and a model is open. Call manage_model_connection with operation=connect.' or 'If field_not_found, use search_model to locate the correct field name and retry.'
Destructive operations (pbir_delete_object, manage_measure DELETE, manage_table DELETE) lack confirmation or dry-run support. Agents can delete pages, measures, or tables without a confirmation step, risking irreversible data loss.
Parameter descriptions lack format and constraint details. Examples: pbir_create_page expects a name but does not specify max length, allowed characters, or uniqueness rules. pbir_update_visual_layout expects x/y/width/height but lacks pixel range constraints (e.g., 1 - 9999).
Generic descriptions fail to distinguish similar tools or clarify when to use them. pbir_inspect_structure ('Debugs the folder structure') does not explain when or why an agent should call it vs pbir_get_info. manage_table and manage_column both modify tables but lack clear separation of concerns.
No tool accepts or documents pagination for list operations. pbir_list_visuals and list_objects likely return entire lists without limit/offset/page_size parameters, risking context window overflow and degraded LLM reasoning on large projects.
Connection handling scattered across tools. manage_model_connection requires explicit 'connect' operation, but pbir_* tools may auto-detect connections. Unclear which tools require prior connection setup or error messaging when connection is missing.
Implement dry-run or confirmation pattern for destructive tools. Add an optional 'confirm' parameter (default=false) to pbir_delete_object and manage_* DELETE operations. When confirm=false, return {status: 'preview', items_to_delete: [...], message: 'Preview only. Call again with confirm=true to execute.'} Example: pbir_delete_object with confirm parameter returns preview of what would be deleted, requiring explicit confirmation in a second call.
Add pagination to list and search tools. Example: pbir_list_visuals(page_name, limit=50, offset=0) returns {visuals: [...], total_count: 150, has_more: true, next_offset: 50}. Limit default to 50 per Arcade baseline to prevent context overflow.
Clarify tool dependencies in descriptions. Add prerequisite hints like: pbir_bind_measure description: 'Requires pbir_create_visual to be called first on the target page. Use pbir_list_visuals to see available visuals before binding.'
Split manage_model_connection into two tools: get_model_connection (returns current connection status) and set_model_connection (establishes new connection). Each with clear, focused responsibility.
Add scope declarations to sensitive tools. Example: manage_role with scope 'admin:security' and manage_table with scope 'admin:model' to signal permission requirements and audit trail expectations.