Repository navigation
DevRay DataLab MCP v0.3.4
DevRay DataLab MCP v0.3.4 adds model selection to leaderboard calls and improves the guidance
available to assistants through all 27 MCP tool descriptions. The server runs over stdio on
Python 3.11 or later and requires no cloud account or API key.
Compare specific models
Call get_model_leaderboard with an existing workspace dataset, its target column, and the task
type. The optional models array accepts unique model IDs for that task. For example, with the
classification sample from this repository:
{
"name": "sample_classification.csv",
"target_column": "purchased",
"task_type": "classification",
"models": ["decision_tree", "logistic_regression"],
"metric": "f1",
"max_trials": 2
}Each call runs fresh cross-validation and returns the ranked trials, the winner and its score,
and whether Optuna ran. Omit models to use the previous default candidates. Ties follow evaluation
order, and the trial and time budgets can limit which candidates are scored. Invalid selections
return validation_error. Use automl_run for full run metadata.
The AutoML tools now honor delimiter for CSV files. Optuna is optional and runs only when
installed and tuning budget remains. Pass a split_manifest_id to reserve the manifest's test
rows for evaluation.
Reliability improvements
- Timestamp split manifests preserve chronological order for CV, including on reuse of legacy manifests.
- Candidate evaluation and Optuna share a monotonic deadline. A trial in progress can finish after
the deadline. - Seeded tuning is reproducible, and the winner comes from the highest-ranked successful trial.
- Cooperative training cancellation records the cancelled terminal state.
- Explicit tuning accepts every supported model ID for the given task.
- Prediction returns
validation_errorfor invalid artifact identifiers.
Installation and Glama
Use Python 3.11+ and launch the installed devray-datalab-mcp command from a stdio client. Set
DEVRAY_WORKSPACE_DIR to the server-side dataset directory and DEVRAY_ARTIFACTS_DIR to a writable
directory. Set DEVRAY_EXPERIMENTS_DB_PATH to a file under a writable directory when the client
launches the process from a read-only working directory.
Glama can build from PyPI or infer a container build and wrap the existing stdio command. Datasets
must be readable in the environment running the server. Hosted verification for this release
candidate is pending. Glama hosting documentation