Exposes a full data-preparation and feature-engineering pipeline as Model Context Protocol (MCP) tools for profiling, planning, cleaning, validating, and ML-preparing datasets without a UI.
DataPrepAgent exposes 7 tools with clear verb-noun naming (profile_data, suggest_cleaning_plan, clean_data, validate_cleaning, list_supported_formats, recommend_feature_engineering, apply_feature_engineering). All tools have descriptions (avg ~150 chars, within 10-1024 range). Input schemas are present with typed parameters and descriptions. However, output schemas are NOT documented, tools return JSON strings but the structure (ProfileReport, CleaningPlan, etc.) is only referenced in docstrings, not formally specified. Parameter descriptions are adequate but lack constraint details (ranges, enums, format specs). Error handling is minimal, no recovery guidance or categorization. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk profiles (READ_ONLY vs WRITE). Missing pagination guidance for list_supported_formats.
Execute an approved feature engineering plan on a cleaned dataset. Only actions with approved=true are executed. Each action is fault-tolerant: failures are logged but do not abort the pipeline.
Execute a cleaning plan (from suggest_cleaning_plan) on a data file. Only actions with approved=true are executed.
Return the list of file formats DataPrepAgent can ingest.
Ingest a CSV, Excel, JSON, or XML file and return a full data-quality profile as a JSON string (ProfileReport schema).
Analyze a cleaned dataset and return an AI-generated feature engineering plan. Performs two-phase analysis: 1. Statistical: skewness, cardinality, correlation, variance analysis 2. LLM reasoning: recommends encoding, scaling, distribution transforms, feature creation, and feature selection actions
Output schemas not formally documented. Tools return JSON strings referencing ProfileReport, CleaningPlan, FeatureEngineeringPlan, etc., but these schemas are not visible in the MCP tool definitions. LLMs cannot infer downstream field names or structure for chaining.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools have explicit risk profiles (READ_ONLY vs WRITE) documented in comments, but not exposed via MCP annotations. Agents cannot distinguish safe vs destructive operations without parsing descriptions.
Parameter constraints not formally specified. 'file_path' and 'target_column' lack format/pattern constraints, length limits, or enum values. Descriptions mention 'absolute or relative path' and 'optional' but do not enforce via JSON Schema.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Given a data file and its ProfileReport JSON (from profile_data), return an AI-generated cleaning plan as a JSON string (CleaningPlan schema).
Compare the original and cleaned files and return a ValidationReport.
No error handling guidance. Tools do not document what errors can occur, how to recover, or what the LLM should do next. E.g., profile_data could fail on unsupported file format, missing file, or parse error, but no recovery hints provided.
list_supported_formats lacks output schema. Tool returns a list but does not document structure (array of strings? objects with format + extension?). No pagination guidance despite potential for future expansion.