DevRay DataLab MCP
A local-first, open-source Model Context Protocol (MCP) server for data science and machine learning. It lets AI assistants (Claude, Cursor, Cline, Kilo Code, and other MCP-compatible clients) perform real exploratory data analysis and train/evaluate machine learning models entirely on the user's own computer -- no cloud services, API keys, or paid subscriptions required.
Status
Working local MCP server over stdio transport. All dataset processing and model training run locally via scikit-learn, pandas, and numpy.
Features
Dataset tools
list_datasets-- list CSV datasets in the configured workspace (filename, format, size).inspect_dataset-- load a CSV and report rows, columns, data types, missing values, duplicate rows, numerical statistics, categorical summaries, and a configurable preview.analyze_correlations-- Pearson/Spearman correlations among numerical variables, strongest positive/negative pairs, constant/unsuitable columns.analyze_data_quality-- missing values, duplicates, constant columns, suspected outliers (IQR), and preprocessing recommendations without modifying the original data.
Machine learning tools
prepare_dataset-- define target + task type; fit scikit-learn preprocessing (imputation, encoding, scaling) on training data only; preserve the fitted transform for prediction.train_baseline-- train logistic regression, decision tree, random forest (classification) or linear/ridge regression, decision tree, random forest (regression); train/test split with optional stratification; evaluation metrics.compare_models-- train and compare multiple algorithms on consistent splits; select the best by a user-specified metric (not accuracy by default for every dataset).export_model-- save a fitted model with its preprocessing pipeline, metadata, feature schema, hyperparameters, metrics, and package versions.predict-- load a previously exported local model, validate features, generate predictions (+ probabilities where supported), handle schema mismatches.
System tools
system_info-- OS, CPU cores, RAM, storage, GPU detection, and whether a GPU training backend is installed. scikit-learn runs on CPU by default.
Supported algorithms
Classification: Logistic Regression, Decision Tree, Random Forest. Regression: Linear Regression, Ridge Regression, Decision Tree, Random Forest.
Architecture
devray-datalab-mcp/
├── src/devray_datalab/
│ ├── server.py # MCP stdio server, tool registry
│ ├── config.py # runtime configuration (env overridable)
│ ├── core/
│ │ ├── loader.py # CSV loading, dataset listing
│ │ ├── preprocessing.py # scikit-learn ColumnTransformer pipelines
│ │ ├── trainer.py # model training, job/resource constraints
│ │ ├── evaluator.py # classification/regression metrics
│ │ └── hardware.py # CPU/RAM/storage/GPU detection (psutil)
│ ├── tools/
│ │ ├── datasets.py # list_datasets, inspect_dataset
│ │ ├── analysis.py # correlations, data quality
│ │ ├── training.py # prepare_dataset, train_baseline, compare_models
│ │ ├── models.py # export_model, predict
│ │ ├── storage.py # joblib artifact persistence
│ │ └── system.py # system_info
│ └── utils/
│ ├── validation.py # path traversal protection, input validation
│ └── errors.py # structured error types
├── datasets/ # configured workspace (datasets)
├── artifacts/ # trained model artifacts
├── examples/ # example datasets
└── tests/ # pytest suite incl. stdio e2e
- Local only: stdio transport; no network, no cloud.
- Security: all file access is confined to the workspace; path traversal is blocked; file extensions/sizes validated; only artifacts created by this app are loaded for prediction.
- Resource management: configurable CPU thread limits, training memory budget, single concurrent training job, training timeouts, cooperative cancellation.
- No GPU deps: PyTorch/TensorFlow/CUDA are not required or installed.
Installation
Prerequisites
Python 3.11+ (tested with 3.15).
Installation methods
Choose one of the following:
Option A: Run directly with uvx (no install)
uvx devray-datalab-mcp
Option B: Install from PyPI into a virtual environment (recommended)
# Windows PowerShell
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install devray-datalab-mcp
devray-datalab-mcp
# Linux / macOS
python3 -m venv .venv
source .venv/bin/activate
pip install devray-datalab-mcp
devray-datalab-mcp
Option C: Install from GitHub source
# Windows PowerShell
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install git+https://github.com/rayklanderman/devray-datalab-mcp.git
devray-datalab-mcp
# Linux / macOS
python3 -m venv .venv
source .venv/bin/activate
pip install git+https://github.com/rayklanderman/devray-datalab-mcp.git
devray-datalab-mcp
Option D: Editable install from a local clone (development)
# Windows PowerShell
git clone https://github.com/rayklanderman/devray-datalab-mcp.git
cd devray-datalab-mcp
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -e .
pip install pytest ruff
devray-datalab-mcp --workspace ./datasets --artifacts ./artifacts
# Linux / macOS
git clone https://github.com/rayklanderman/devray-datalab-mcp.git
cd devray-datalab-mcp
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
pip install pytest ruff
devray-datalab-mcp --workspace ./datasets --artifacts ./artifacts
Exact Windows PowerShell installation commands
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install devray-datalab-mcp
devray-datalab-mcp
Configuration via environment variables (see .env.example):
| Variable | Default | Meaning |
|---|---|---|
DEVRAY_WORKSPACE_DIR |
datasets |
Dataset workspace directory |
DEVRAY_ARTIFACTS_DIR |
artifacts |
Model artifacts directory |
DEVRAY_MAX_DATASET_SIZE_BYTES |
209715200 (200 MB) |
Max dataset upload size |
DEVRAY_MAX_PREVIEW_ROWS |
25 |
Max preview rows |
DEVRAY_MAX_CPU_THREADS |
(unset) | CPU thread limit for training |
DEVRAY_TRAINING_MEMORY_BUDGET_MB |
(unset) | Training memory budget |
DEVRAY_MAX_CONCURRENT_JOBS |
1 |
Max concurrent training jobs |
DEVRAY_TRAINING_TIMEOUT_SECONDS |
1800 |
Training timeout |
DEVRAY_LOG_LEVEL |
INFO |
Log level (logs to stderr) |
MCP client setup
The server speaks MCP over stdio. Logging/diagnostics go to stderr, so stdout is reserved for MCP protocol messages.
Example client configuration
The server speaks MCP over stdio. Logging/diagnostics go to stderr, so stdout
is reserved for MCP protocol messages. Replace C:\path\to\project with your
actual project directory, or use the published devray-datalab-mcp command
when installed from PyPI.
Kilo Code
{
"mcpServers": {
"devray-datalab": {
"command": "C:\\path\\to\\project\\.venv\\Scripts\\python.exe",
"args": ["-m", "devray_datalab.server"],
"cwd": "C:\\path\\to\\project",
"env": {
"DEVRAY_WORKSPACE_DIR": "C:\\path\\to\\project\\datasets",
"DEVRAY_ARTIFACTS_DIR": "C:\\path\\to\\project\\artifacts"
}
}
}
}
Cline / Cursor / Claude Desktop
These clients use the same mcpServers JSON structure:
{
"mcpServers": {
"devray-datalab": {
"command": "devray-datalab-mcp",
"args": [],
"env": {
"DEVRAY_WORKSPACE_DIR": "C:\\path\\to\\project\\datasets",
"DEVRAY_ARTIFACTS_DIR": "C:\\path\\to\\project\\artifacts"
}
}
}
}
When using the published devray-datalab-mcp command (installed from PyPI or
uvx), omit cwd and let the client resolve the executable from PATH. Set
DEVRAY_WORKSPACE_DIR/DEVRAY_ARTIFACTS_DIR to absolute paths outside the
project if you prefer dedicated dataset/artifact directories.
VS Code MCP-compatible tooling
Add the server to your VS Code MCP settings (settings.json):
{
"mcp.servers": {
"devray-datalab": {
"command": "devray-datalab-mcp",
"args": [],
"env": {
"DEVRAY_WORKSPACE_DIR": "${workspaceFolder}/datasets",
"DEVRAY_ARTIFACTS_DIR": "${workspaceFolder}/artifacts"
}
}
}
}
MCP Inspector
The official MCP Inspector can be used to test the server:
npx @modelcontextprotocol/inspector
# Connect to: devray-datalab-mcp (stdio)
# Command: devray-datalab-mcp
Dataset workspace configuration
Designate authorized local directories for datasets and model artifacts using environment variables:
DEVRAY_WORKSPACE_DIR-- directory containing CSV datasets (default:datasets).DEVRAY_ARTIFACTS_DIR-- directory where trained model artifacts are stored (default:artifacts).
Both directories are created automatically if missing. All file access is confined to these directories; path traversal is blocked. Point them at absolute paths outside your project if you want a shared workspace across clients.
Example (PowerShell):
$env:DEVRAY_WORKSPACE_DIR="C:\ml-data\datasets"
$env:DEVRAY_ARTIFACTS_DIR="C:\ml-data\artifacts"
devray-datalab-mcp
Available tools and example arguments
| Tool | Example arguments |
|---|---|
list_datasets |
{} |
inspect_dataset |
{"name": "sample_classification.csv", "max_rows": 10} |
analyze_correlations |
{"name": "sample_classification.csv", "method": "pearson", "top_n": 5} |
analyze_data_quality |
{"name": "sample_classification.csv"} |
prepare_dataset |
{"name": "sample_classification.csv", "target_column": "purchased", "task_type": "classification"} |
train_baseline |
{"name": "sample_classification.csv", "target_column": "purchased", "task_type": "classification", "model_name": "random_forest", "random_state": 42} |
compare_models |
{"name": "sample_classification.csv", "target_column": "purchased", "task_type": "classification", "models": ["logistic_regression", "random_forest"], "metric": "f1"} |
export_model |
{"artifact_name": "my_model", "name": "sample_classification.csv", "target_column": "purchased", "task_type": "classification", "model_name": "random_forest"} |
predict |
{"artifact_name": "my_model", "features": {"age": 40, "income": 80000, "credit_score": 720, "category": "electronics", "region": "east"}} |
system_info |
{} |
Example machine learning workflow
Using sample_classification.csv (target purchased, binary classification):
-
Inspect the dataset:
inspect_datasetwith{"name": "sample_classification.csv"}. -
Check data quality:
analyze_data_qualitywith{"name": "sample_classification.csv"}. -
Check correlations:
analyze_correlationswith{"name": "sample_classification.csv", "method": "pearson"}. -
Train & compare models:
compare_modelswith{"name": "sample_classification.csv", "target_column": "purchased", "task_type": "classification", "metric": "f1"}. -
Export the best model:
export_modelwith{"artifact_name": "churn_model", "name": "sample_classification.csv", "target_column": "purchased", "task_type": "classification", "model_name": "random_forest"}. -
Predict on new data:
predictwith{"artifact_name": "churn_model", "features": {...}}.
Example regression workflow
Using sample_regression.csv (target spend, numeric target):
- Inspect the dataset:
inspect_datasetwith{"name": "sample_regression.csv"}. - Check data quality:
analyze_data_qualitywith{"name": "sample_regression.csv"}. - Check correlations:
analyze_correlationswith{"name": "sample_regression.csv", "method": "pearson"}. - Train & compare models:
compare_modelswith{"name": "sample_regression.csv", "target_column": "spend", "task_type": "regression", "metric": "r2"}. - Export the best model:
export_modelwith{"artifact_name": "spend_model", "name": "sample_regression.csv", "target_column": "spend", "task_type": "regression", "model_name": "random_forest"}. - Predict spend on new data:
predictwith{"artifact_name": "spend_model", "features": {...}}.
Release links
- GitHub repository: https://github.com/rayklanderman/devray-datalab-mcp
- GitHub releases: https://github.com/rayklanderman/devray-datalab-mcp/releases
- PyPI package: https://pypi.org/project/devray-datalab-mcp/
Example prompts
- "Inspect the sample_classification.csv dataset and tell me about missing values and column types."
- "Analyze correlations in sample_regression.csv between income and spend."
- "Train a random forest to predict 'purchased' from sample_classification.csv and compare it against logistic regression using F1."
- "Export the trained model as 'churn_model' and predict purchased for a 40-year-old in the east region with income 80000 and credit score 720."
- "Show me the CPU, RAM, and storage available for local training."
Hardware requirements
- Designed for modest hardware (older Windows laptops with 8-16 GB RAM).
- CPU training is the default; scikit-learn does not use GPU acceleration by default.
- No PyTorch, TensorFlow, or CUDA required.
- Dataset size limit defaults to 200 MB.
Security and privacy
- All processing is local; no data leaves your machine.
- Datasets and artifacts are confined to the configured workspace/artifacts directories.
- Path traversal and unauthorized filesystem access are blocked.
- Only CSV datasets are accepted; file size and extensions are validated.
- No arbitrary code or shell execution; no external downloads; no uploads to third parties.
- Uploaded datasets and model artifacts are treated as untrusted; only artifacts created and validated by this application are loaded for prediction.
- Original datasets are never overwritten without explicit permission.
Limitations
- Single concurrent training job by default.
- scikit-learn models run on CPU; no GPU acceleration in this release.
- Classification stratification requires at least 2 samples per class.
- ROC-AUC is reported only for binary classification.
- CSV is the only supported dataset format.
- No built-in job queue/status polling tool; training runs within the tool call with a configurable timeout.
- Cancellation is cooperative (checked during iterative training loops).
Roadmap
- Optional GPU-accelerated backend plug-in (without redesigning MCP tools).
- Training job status/polling and cancellation via MCP progress mechanisms.
- Additional models (SVM, k-NN, gradient boosting).
- More preprocessing options and feature selection.
- Dataset profiling and automated preprocessing recommendations.
Versioning and release notes
The project follows semantic versioning. The current
release is v0.1.0. Releases are published to PyPI as
devray-datalab-mcp and registered in the MCP Registry as
io.github.rayklanderman/devray-datalab-mcp.
v0.1.0 (initial release)
- Ten MCP tools: list_datasets, inspect_dataset, analyze_correlations, analyze_data_quality, prepare_dataset, train_baseline, compare_models, export_model, predict, system_info.
- Local-first stdio transport; scikit-learn CPU training.
- Path-traversal protection, dataset size/extension validation.
- 31 automated tests including a stdio end-to-end client test.
- GitHub Actions CI/CD and PyPI Trusted Publishing workflows.
Contributing
- Fork the repository.
- Create a feature branch (
git checkout -b feature/...). - Install dev dependencies:
pip install -e . pytest ruff. - Run tests:
pytest(or.venv\Scripts\python.exe -m pyteston Windows). - Run lint:
ruff check src/ tests/andruff format src/ tests/. - Commit with a concise message and open a pull request.
License
MIT License. See LICENSE.
Metadata
Release files for devray-datalab-mcp 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| devray_datalab_mcp-0.1.0.tar.gz | 36.9 kB | With provenance | Details for devray_datalab_mcp-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| devray_datalab_mcp-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for devray_datalab_mcp-0.1.0-py3-none-any.whl |
Total release size: 73.0 kB
Release files / devray_datalab_mcp-0.1.0.tar.gz
| Download URL | devray_datalab_mcp-0.1.0.tar.gz |
|---|---|
| Size | 36.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
6e564dab73f2da79583597e7dd69c45ee7a91d496bf2fb9f4e15b8bc3531ef2a
|
|
BLAKE2b-256 checksum How to use checksums |
74e0a3417c47783bd5e63465aab73694705a005513d40c9e4955d172da14df77
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.
Transparency logRelease files / devray_datalab_mcp-0.1.0-py3-none-any.whl
| Download URL | devray_datalab_mcp-0.1.0-py3-none-any.whl |
|---|---|
| Size | 36.1 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
a6275485c65dd3998bf75d0c617b9f3934e12688aed625aa5dd57da327be01a7
|
|
BLAKE2b-256 checksum How to use checksums |
71a54033863949ede4e8e26845e614bc50d79637c79429caed48db99da2e69a1
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.
Transparency logSecurity for devray-datalab-mcp 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for devray_datalab_mcp-0.1.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.