Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
The MCP Excel SQL Server has 7 tools with clear naming conventions and mostly adequate schemas. However, there are significant gaps in parameter descriptions, missing output schema documentation, and incomplete error handling guidance. Tool names follow verb_noun patterns appropriately (load_dir, query_excel, create_view, list_tables, get_table_info, list_views, drop_view). Descriptions exist for all tools but vary in quality and specificity. Input schemas are present and include type information, but many parameters lack detailed descriptions explaining constraints, formats, and expected values. Output schemas are not documented in the visible code, making it unclear what data structures agents should expect. Error handling is minimal, the code shows validation but no recovery guidance in tool descriptions.
Tools (7)
create_viewwritesource verified73/100
Create a named SQL view for reuse
drop_viewreversiblesource verified70/100
Remove a view
get_table_inforead only50/100
Get detailed information about a specific table including schema
list_tablesread onlysource verified83/100
List all available tables from loaded Excel files
list_viewsread only50/100
List all available views
load_dirwritesource verified73/100
Load a directory of Excel files and make them queryable via SQL
Output schemas not documented. The visible code does not show tool.outputSchema definitions, so agents cannot predict what fields will be returned. For example, query_excel likely returns results but the structure is opaque to the LLM.
Parameter descriptions lack format, range, and constraint details. The 'path' parameter in load_dir says 'Path to directory...' but does not explain: must be absolute or relative? What happens if symlinks exist? The 'include_glob' and 'exclude_glob' parameters lack examples and specifics on glob syntax accepted.
No error recovery guidance in tool descriptions. The code shows validation (e.g., _validate_root_path, _validate_view_name) but tool descriptions do not state what errors are possible or how the LLM should respond. For example, create_view does not say 'If a view with this name exists, it will be replaced' or what SQL syntax errors look like.
create_view
Recommendations
Document output schemas for all tools. Specify what query_excel returns (e.g., { rows: Array<Record<string, any>>, rowCount: number, columns: Array<{name: string, type: string}> }). Same for get_table_info, list_tables, list_views. This allows agents to plan downstream uses of returned data.
Enhance parameter descriptions with format constraints. Update 'path' in load_dir: 'Absolute or relative path to a directory. Symlinks are followed. Must be readable by the server process.' Update 'sql' in query_excel: 'SQL SELECT query using DuckDB dialect. Supports CTEs, subqueries, aggregations. Timeouts after 30s. UPDATE/DELETE/INSERT operations are not supported.'
Add error recovery hints to each tool. Example for query_excel: 'If your SQL has syntax errors, the tool returns the DuckDB error message. Try simplifying the query or call get_table_info(table_name) first to verify table and column names.' Example for create_view: 'View names must be unique and contain only alphanumeric characters and underscores. If a view already exists, it will be replaced (no confirmation).'
Implement confirmation flow for destructive operations. Modify drop_view description: 'Permanently removes a named view. This action cannot be undone. No user confirmation is currently required, ensure you intend to delete before calling.' Consider adding a confirm_before_delete parameter (boolean, defaults to false) and update description to explain confirmation behavior.
Define the 'overrides' schema explicitly. Replace generic 'object' with a schema listing allowed keys: { skip_rows: number, header_rows: number, skip_footer: number, range: string, drop_regex: string, column_renames: object }. Or use a discriminated union: { sheet_name: string, overrides: { type: 'skip_rows'|'header_rows'|..., value: any } }.
Spec posture evidence
Inferred effective spec: <=2025-11-25.
Relies on Logging (deprecated) - log to stderr or use OpenTelemetry
Irreversible operations (drop_view, create_view with overwrite) lack confirmation or dry-run support. An agent could accidentally drop a view without undo capability. The code supports view deletion but tool description does not warn of irreversibility.
Tool 'query_excel' accepts free-form 'sql' parameter with no validation hints. Agents may pass invalid SQL or dangerous operations. Description does not state: What SQL dialect is supported? Can agents use UNION, subqueries, CTEs? Can they INSERT/UPDATE/DELETE? Are there query timeouts?
The 'overrides' parameter in load_dir accepts a generic object without schema constraints. Agents cannot know valid override keys (skip_rows, header_rows, etc.) without reading comments or external docs. Should be an explicit schema or enum of known override patterns.
load_dir
Document SQL dialect and limitations in query_excel. Add: 'Uses DuckDB SQL syntax. Supports: SELECT, WITH (CTEs), JOIN, GROUP BY, ORDER BY, LIMIT, aggregates (COUNT, SUM, AVG, MAX, MIN). Not supported: transactions, stored procedures, multi-statement queries, DDL (CREATE TABLE). Result set capped at 10,000 rows.'
Add pagination to list_tables and list_views if they can return large result sets. Include limit and offset parameters, and return { items: [...], total: number, limit: number, offset: number }. Current no-param design could cause context overflow if many tables/views exist.
Specify valid values for glob patterns in load_dir. Example: 'include_glob: Glob patterns (e.g., ["**/*.xlsx", "**/*.csv"]). If omitted, defaults to ["**/*.xlsx", "**/*.xlsm", "**/*.xls", "**/*.csv", "**/*.tsv"]. exclude_glob: Patterns to skip (e.g., ["**/archive/**", "**/*backup*"]).'
Return related IDs to enable tool chaining. When get_table_info returns table metadata, also return an available_views field listing views that reference this table (if applicable). This helps agents discover related data without extra lookup calls.