MCP Tool Reference

Complete reference for every tool the sktime-mcp server exposes to an MCP client. The server currently registers 26 tools.

You normally do not call these by hand — the assistant selects them from your natural-language request. This page exists so you can see exactly what the assistant has available, what each argument means, and what comes back.

For the Python API of the package itself (classes, modules), see the API Reference page instead.

---

Conventions

Handles

Most tools return or accept a handle — a string ID naming an object the server holds in memory.

Prefix	Created by	Refers to
est_…	instantiate, load_model	An estimator or pipeline
data_…	load_data_source, split_data, transform_data	A time series (target y plus any exogenous X)

Handles live in the server process and are lost when it restarts. Use save_data, save_model, or export_code to persist anything you need to keep, and release_handle / release_data_handle to free memory.

Asynchronous execution: the run_async flag

Long-running work is made asynchronous by passing run_async: true to the tool itself. There are no separate *_async tools.

Exactly four tools accept the flag:

• fit

• predict

• evaluate

• load_data_source

Behaviour:

run_async	Returns	Use when
false (default)	The real result — a handle, predictions, scores	The call is quick
true	{"success": true, "job_id": "..."} immediately	Training or loading is slow

When you pass run_async: true, the work is scheduled on the server’s event loop and you get a job_id back straight away. Track it with the job tools:

fit(run_async=true)  ->  job_id
        |
        +-- check_job_status(job_id)   poll status and progress
        +-- list_jobs()                see everything in flight
        +-- cancel_job(job_id)         stop it

Note

There is no push notification when a job finishes. The client must poll check_job_status to find out. An assistant will typically do this for you when you ask “is it done yet?”, but nothing arrives unprompted.

Results are retrieved through check_job_status once status is completed.

---

Discovery

query_registry

Discover sktime estimators, metrics, or capability tags.

Argument	Type	Required	Default	Description
task	string	—	—	Filter by scitype: forecaster, classifier, regressor, transformer, clusterer, detector, splitter, metric, param_est, aligner, network. Set to tag/tags to list available tags instead.
tags	object	—	—	Filter by capability tags, e.g. {"capability:pred_int": true}. Ignored when task="tag".
query	string	—	—	Case-insensitive substring search over name and description. Combines with task and tags.
limit	integer	—	50	Maximum results.
offset	integer	—	0	Skip this many results (pagination).

Commonly useful tags: capability:pred_int (prediction intervals), capability:multivariate, handles-missing-data, scitype:y.

describe_component

Detailed information about any class in the sktime ecosystem — estimators, splitters, metrics, transformers.

Argument	Type	Required	Description
name	string	✅	Class name, e.g. ARIMA, SlidingWindowSplitter, MeanAbsolutePercentageError.

list_available_data

Lists demo datasets and active user-loaded data handles in one response.

Argument	Type	Required	Description
is_demo	boolean	—	true = demos only, false = live handles only, omit = both.

---

Instantiation and handles

instantiate

Create an estimator or pipeline from an sktime craft specification string. Composition is expressed directly in the spec — sktime validates it.

Argument	Type	Required	Description
spec	string	✅	Craft spec, e.g. ARIMA(order=(1, 1, 1)) or Detrender() * ARIMA().

Returns an est_… handle.

list_handles

Lists all active estimator handles. Takes no arguments.

release_handle

Argument	Type	Required	Description
handle	string	✅	Estimator handle to release.

---

Execution

fit

Fit an estimator. Supply X/y as either live handles or demo dataset names, depending on the estimator’s scitype.

Argument	Type	Required	Default	Description
estimator_handle	string	✅	—	Handle from instantiate.
X_handle	string	—	—	Data handle for X (features, panel).
y_handle	string	—	—	Data handle for y (target, labels).
X_dataset	string	—	—	Demo dataset name for X.
y_dataset	string	—	—	Demo dataset name for y.
fh	int or list	—	—	Forecast horizon passed through to fit.
run_async	boolean	—	false	Run in the background, return a job_id. See run_async.

predict

Generate predictions from a fitted estimator.

Argument	Type	Required	Default	Description
estimator_handle	string	✅	—	Handle of a fitted estimator.
horizon	integer	—	12	Forecast horizon.
mode	string	—	predict	One of predict, predict_interval, predict_quantiles, predict_proba, predict_var.
coverage	float or list	—	0.9	Coverage level(s) — used by predict_interval.
alpha	float or list	—	—	Quantile level(s) — used by predict_quantiles.
X_handle	string	—	—	Data handle for X.
y_handle	string	—	—	Data handle for y (needed by detectors/annotators).
X_dataset	string	—	—	Demo dataset name for X.
y_dataset	string	—	—	Demo dataset name for y.
run_async	boolean	—	false	Run in the background, return a job_id.

Interval and quantile forecasts are modes of this tool, not separate tools.

update

Update a fitted estimator with new data.

Argument	Type	Required	Description
estimator_handle	string	✅	Handle of a fitted estimator.
X_handle / y_handle	string	—	Data handles for the new data.
X_dataset / y_dataset	string	—	Demo dataset names for the new data.

get_fitted_params

Argument	Type	Required	Description
estimator_handle	string	✅	Handle of a fitted estimator.

call_method

Escape hatch: call any native method on an instantiated component. Use this for scitypes that do not fit the fit/predict shape — splitters, metrics, aligners.

Argument	Type	Required	Description
handle_id	string	✅	Handle of the instantiated component.
method_name	string	✅	Method to call, e.g. split, get_alignment, __call__.
kwargs	object	—	Keyword arguments. Keys suffixed _dataset or _data_handle inject server-side data, e.g. {"y_dataset": "airline"}.

evaluate

Cross-validate an estimator.

Argument	Type	Required	Default	Description
estimator_handle	string	✅	—	Handle from instantiate.
y	string	✅	—	Target series: a data handle ID or a demo dataset name.
X	string	—	—	Exogenous series: data handle ID or demo dataset name.
cv_folds	integer	—	3	Number of folds. Ignored when initial_window is set.
metric	string	—	—	Metric name, e.g. MeanAbsolutePercentageError.
initial_window	integer	—	—	Initial training window for expanding-window CV.
run_async	boolean	—	false	Run in the background, return a job_id.

---

Data

load_data_source

Load data into a data_… handle. The config object must carry a type key.

Argument	Type	Required	Default	Description
config	object	✅	—	Source configuration; must include type.
run_async	boolean	—	false	Load in the background, return a job_id.

Supported config.type values:

type	Additional keys
pandas	data, time_column, target_column
file	path (CSV / .xlsx / Parquet), time_column, target_column
sql	connection and query keys, time_column, target_column
url	path/URL, time_column, target_column

{
  "type": "file",
  "path": "/data/sales.csv",
  "time_column": "date",
  "target_column": "revenue"
}

inspect_data

Rich metadata for a loaded handle: mtype, scitype, shape, column names, dtypes, index level names, inferred frequency, cutoff, missing-value count, a 5-row head preview, and per-column summary statistics.

Argument	Type	Required	Description
data_handle	string	✅	Handle to inspect.

split_data

Split a series into temporal train/test handles. Provide exactly one of test_size or fh.

Argument	Type	Required	Description
data_handle	string	✅	Handle to split.
test_size	number	—	Hold-out fraction, exclusive range (0.0, 1.0). Mutually exclusive with fh.
fh	int or list	—	Integer: hold out that many final steps. List: holds out max(fh) steps — fh=[1,5,10] reserves 10. Mutually exclusive with test_size.

Returns train_handle and test_handle.

transform_data

Returns a new handle; the input handle is unchanged.

Argument	Type	Required	Default	Description
data_handle	string	✅	—	Handle to transform.
action	string	—	format	format or convert.
auto_infer_freq	boolean	—	true	(format) Infer and set DatetimeIndex frequency.
fill_missing	boolean	—	true	(format) Forward/backward-fill missing values.
remove_duplicates	boolean	—	true	(format) Drop duplicate timestamps, keeping the first.
to_mtype	string	—	—	(convert, required) Target mtype, e.g. pd.DataFrame, pd.Series, np.ndarray.

action="format" also fills index gaps and reports what it did in changes_applied.

save_data

Writes the target y and any exogenous X behind a handle to one file, creating parent directories as needed.

Argument	Type	Required	Default	Description
data_handle	string	✅	—	Handle to export.
path	string	✅	—	Destination path.
format	string	—	csv	csv, parquet, or json.

Warning

The format comes from the format argument, not the file extension. Writing to out.parquet with the default format produces a CSV.

release_data_handle

Argument	Type	Required	Description
data_handle	string	✅	Data handle to release.

plot_series

Plot one or more series. Saves a file when path is given, otherwise returns the image as a base64 string.

Argument	Type	Required	Default	Description
data_handles	array of string	✅	—	Handles to plot, e.g. train, test, forecast.
labels	array of string	—	—	Legend label per handle.
title	string	—	—	Plot title.
path	string	—	—	Save location, e.g. /tmp/plot.png. Omit to get base64 back.
figsize	array	—	[12, 6]	[width, height] in inches.
dpi	integer	—	150	Resolution.
markers	string or array	—	—	Marker style(s), e.g. "o" or [".", "x"].
x_label / y_label	string	—	—	Axis labels.
image_format	string	—	png	png, svg, or webp.

---

Persistence and code generation

export_code

Emit standalone, runnable Python that reconstructs the estimator.

Argument	Type	Required	Default	Description
handle	string	✅	—	Estimator/pipeline handle.
var_name	string	—	model	Variable name in the generated code.
include_fit_example	boolean	—	false	Append a fit/predict example.
dataset	string	—	airline	Dataset used in the example.

save_model

Argument	Type	Required	Description
estimator_handle	string	✅	Estimator to save.
path	string	✅	Local directory or URI.
mlflow_params	object	—	Extra parameters for sktime.utils.mlflow_sktime.save_model.

load_model

Argument	Type	Required	Description
path	string	✅	Path to the saved model directory.

Registers the loaded model and returns a new est_… handle.

---

Background jobs

See the run_async flag for how jobs are created.

check_job_status

Argument	Type	Required	Description
job_id	string	✅	Job to check.

Returns status, progress, and — once completed — the result.

list_jobs

Argument	Type	Required	Default	Description
status	string	—	—	Filter: pending, running, completed, failed, cancelled.
limit	integer	—	20	Page size.
offset	integer	—	0	Records to skip.

Returns the page plus total and has_more for pagination.

cancel_job

Argument	Type	Required	Default	Description
job_id	string	✅	—	Job to cancel.
delete	boolean	—	false	Also remove the job record — useful for tidying completed or failed jobs.

Cancellation is cooperative: the server holds a reference to the running task and cancels it, but work already inside a blocking sktime call finishes its current step before stopping.

---

System

run_command

Danger

This tool executes arbitrary shell commands with the full privileges of the user account running the server. It is not sandboxed, and it is not restricted to the sktime registry.

A client that is allowed to call run_command can read, modify, or delete any file the server process can reach, and can make network requests.

It exists so that an assistant can repair its own environment mid-session — typically installing an optional dependency that an estimator turned out to need:

pip install mlflow

Argument	Type	Required	Description
command	string	✅	The shell command to run.

Treat this as opt-in. Most MCP clients prompt before each tool call and let you allow or deny individual tools — leave run_command on ask-every-time, or disable it outright, unless you specifically want the assistant installing packages and inspecting the filesystem on your behalf. If you need a hard guarantee, run the server inside a container (see the Docker option in the User Guide: Conversational Time-Series Workflows) so the blast radius is the container rather than your machine.

See Architecture for the full trust boundary.