A Model Context Protocol server for managing machine learning training workflows, including dataset management, model training, evaluation, and MCP tool orchestration
7 tools with basic naming and minimal descriptions. All tools have input schemas with type definitions (integer dataset_id/job_id), but descriptions are generic and under 50 chars. No parameter descriptions visible in schema. No output schemas documented. No error handling guidance. Tool names follow verb_noun pattern (inspect, validate, create, get, generate) but lack context about when to use each. Missing composition guidance, unclear how tools chain together or what fields downstream tools expect.
Inspect dataset properties and sample count
Validate whether a dataset is suitable for training and assess risk level
Inspect evaluation metrics for a training job
Generate a data quality and analysis report for a dataset
Generate an evaluation report for a training job
Create a new training job for a dataset
Get the current status of a training job
All tool descriptions under 50 chars and lack WHEN/WHY context. 'Inspect dataset properties and sample count' does not explain when to call this vs validate_trainability, or what the LLM should do with the result.
No output schemas documented. LLMs cannot plan downstream calls or extract required fields. E.g., does dataset.inspect return sample_count, data_quality_score, or both? Does training.create_job return job_id for use in get_status?
Input parameter descriptions missing. Schema shows 'dataset_id' with type integer but no description of valid range, format, or how to obtain it. LLMs cannot infer whether to pass 1 or 'my-dataset'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2025-06-18+ | v2 |
No error handling guidance. If dataset_id is invalid, what error is returned? Can the LLM retry, or should it ask the user? No recovery paths documented.
Tool composition unclear. No documentation of which tools chain together (e.g., does create_job require validate_trainability first?). No indication of required fields in responses for downstream calls.