MCP (Model Context Protocol) layer for sktime - Registry-Driven for LLMs
The sktime-mcp server has 28 tools with significant definition gaps across naming, descriptions, parameter schemas, and output documentation. While tool names generally start with action verbs (query_, describe_, load_, save_, etc.), most lack the depth and specificity required for reliable LLM selection. Critical issues: (1) 16 of 28 tools have descriptions under 100 characters or are truncated/incomplete (e.g., 'instantiate' description cuts off mid-sentence at 'The spec is a string that eva'); (2) Input schemas are visible for some tools (query_registry, describe_component, load_data_source) but many tools (fit, predict, update, evaluate, etc.) have NO visible input schema at all in the source code provided, forcing a schema score of 0 for those tools; (3) Output schemas are not documented for any tool; (4) Parameter descriptions are sparse, most parameters lack guidance on format, range, or dependencies; (5) Error handling and recovery guidance is absent; (6) No tool annotations (readOnlyHint, destructiveHint) despite many tools being destructive (save_model, save_data, delete operations). The server demonstrates domain knowledge (sktime time-series forecasting) but fails to meet production standards for LLM-safe tool definitions. Average per-tool score: 38/100.
Enable/disable automatic formatting when loading data. When enabled, all data loaded via load_data_source will be automatically formatted to be sktime-compatible.
Cancel a running/pending job, and optionally remove its record.
Check the status of a background job.
Get detailed information about ANY class or component in the sktime ecosystem (estimators, splitters, metrics, transformers)
Evaluate an estimator on time series data.
Export Python code for an estimator or pipeline.
Fit an estimator on time series data.
16 tools (fit, predict, update, evaluate, inspect_data, list_handles, release_handle, load_model, list_available_data, plot_series, run_command, save_data, save_model, split_data, transform_data, export_code) have NO visible input schema in the provided source code.
Tool 'instantiate' has description that cuts off mid-sentence ('The spec is a string that eva...'). No input schema visible.
No output/return schemas documented for any tool. LLMs cannot plan downstream calls or know what fields to extract without documented return types. Violates pattern:response-shaper.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | <=2025-11-25 | v2 |
Automatically format time series data to be sktime-compatible. This tool fixes common issues: Missing frequency on DatetimeIndex, Duplicate timestamps, Missing values, Irregular time intervals
Get parameters of a fitted estimator.
Inspect and retrieve statistics about loaded data.
Create an estimator or pipeline instance using a sktime craft specification. The spec is a string that eva
List available built-in datasets for testing and examples.
List all available data source types.
List all active estimator and data handles.
List background jobs with offset/limit pagination.
Load data from any source (pandas, SQL, file, etc.). Can run synchronously (blocking) or asynchronously in the background.
Load a saved sktime model from disk.
Plot time series data and save to PNG.
Make predictions using a fitted estimator.
Discover sktime estimators, metrics, or capability tags. Common tags you can filter estimators by: 'capability:pred_int' (bool) - prediction intervals, 'capability:multivariate' (bool) - multivariate support, 'handles-missing-data' (bool) - NaN handling, 'scitype:y' (str) - target type ('univariate'/'multivariate'/'both'), 'requires-fh-in-fit' (bool) - needs forecast horizon at fit time. Set task='tag' (or 'tags') to query the full list of capability tags.
Release a data handle and free memory.
Release an estimator handle and free memory.
Execute arbitrary Python code in a restricted, safe sandbox.
Save time series data to disk in various formats.
Save a fitted sktime estimator to disk.
Split time series data for train/test evaluation.
Transform time series data using a fitted or unfitted transformer.
Update a fitted estimator with new data.
16 tools with WRITE risk (save_model, save_data, split_data, delete_handle, run_command, etc.) lack tool annotations (destructiveHint, idempotentHint). LLMs cannot distinguish safe from dangerous operations without these hints.
Multiple tools lack error handling and recovery guidance. E.g., 'load_data_source' with type='sql' could fail due to bad credentials, missing table, or network timeout, no error messages tell LLM what to do next. Violates pattern:recovery-guide.
Tool 'run_command' accepts arbitrary Python code with no visible input validation or sandboxing constraints documented. Allows prompt injection and code execution risks. Violates pattern:tool-gateway and pattern:secret-injection.
Parameter descriptions are sparse or missing context. Should use enum constraints instead.
No pagination guidance for list tools. 'list_jobs', 'list_handles', 'list_available_data' accept limit/offset but return behavior is undocumented. Should clarify total count, next_cursor, or whether results are truncated.
Tools like 'release_data_handle', 'release_handle', 'cancel_job' lack clarity on whether these are idempotent. Can calling release_data_handle twice with the same handle_id cause errors or is it safe to retry?
No audit logging or permission declarations visible. Tools like 'save_model', 'run_command', 'delete_handle' should declare required scopes (e.g., 'write:models', 'execute:python'). Violates pattern:scope-declaration and pattern:audit-trail.