Data Science Workspace MCP Server

A local-first Model Context Protocol (MCP) server for data science workspaces.

ds-workspace-mcp gives MCP-compatible assistants structured access to analytical data without handing them arbitrary filesystem or database control. It exposes a guarded local workspace where an assistant can discover datasets, inspect schemas, profile columns, run bounded read-only queries, evaluate baseline modeling readiness, and create reusable markdown modeling reports.

The project is built as a practical reference implementation for AI-assisted data tooling: typed Python, bounded outputs, path traversal protection, read-only SQL guardrails, release metadata, Docker support, and CI-backed documentation.

At a glance

Area	What this server provides
Dataset access	Safe reads for CSV, Parquet, JSON, Excel, SQLite, and DuckDB-backed CSV queries
Exploration	Preview, profiling, missingness summaries, categorical summaries, correlations, and issue detection
Modeling readiness	Target suggestions, feature review, leakage checks, time-series validation, forecast baselines, and dummy baselines
Report workflow	Generate, save, search, preview, compare, copy, rename, delete, and extract markdown report sections
MCP integration	Resources, tools, prompts, stdio transport, Streamable HTTP transport, and runnable client examples
Safety controls	Configured data/report roots, path traversal rejection, size limits, row limits, query length limits, and SQL timeouts

Current scope

Today the repository covers five practical workflows:

• Safe local dataset access for CSV, Parquet, JSON, Excel, and SQLite sources.
• Structured aggregations for common EDA without free-form SQL.
• Lightweight analytical diagnostics, including leakage review, correlations, and time-series checks.
• Baseline modeling planning and markdown report generation.
• Saved report lifecycle management, including cataloging, search, section extraction, section export, and section-aware diffs.

---

Why this project exists

Many AI assistants can reason well, but useful analytical work needs controlled access to real data and real tools.

This MCP server exposes a local analytical workspace without giving the model arbitrary file-system access. The server only reads supported dataset files from a configured data directory, and saved modeling reports live under a separate configured reports directory.

Good use cases:

• connecting an MCP client to a local data science workspace;
• inspecting unfamiliar datasets before deeper analysis;
• creating first-pass EDA, readiness, and baseline modeling reports;
• preserving reusable markdown report artifacts from assistant-led analysis;
• demonstrating AI tooling, data engineering, and security-minded local service design.

Quick start

Install dependencies:

poetry install

Generate a sample dataset:

poetry run generate-sample-healthcare-data

Start the MCP server over Streamable HTTP:

MCP_TRANSPORT=streamable-http poetry run ds-workspace-mcp serve

Open the MCP Inspector and connect to:

http://localhost:8000/mcp

Useful first tool calls:

• summarize_dataset for a compact dataset overview.
• profile_csv for detailed column summaries, distributions, and bounded quality diagnostics.
• aggregate_dataset for allowlisted grouped metrics.
• assess_modeling_readiness for target, feature, leakage, and next-step guidance.
• save_modeling_report for a reusable markdown report artifact.

---

Project structure

ds-workspace-mcp/
├── data/
├── docs/
├── examples/
├── notebooks/
├── src/
│   └── ds_workspace_mcp/
│       ├── cli.py
│       ├── config.py
│       ├── core.py
│       ├── datasets/
│       ├── diagnostics.py
│       ├── experiment_plan.py
│       ├── feature_selection.py
│       ├── modeling_readiness.py
│       ├── modeling_report.py
│       ├── overview.py
│       ├── profiling.py
│       ├── report_export.py
│       ├── reports/
│       ├── server.py
│       ├── timeseries.py
│       ├── ml/
│       ├── sql/
│       └── synthetic/
├── tests/
├── .github/
├── docker-compose.yml
├── Dockerfile
├── README.md
└── pyproject.toml

---

Requirements

• Python 3.11+
• Poetry
• openpyxl is installed as a runtime dependency for .xlsx ingestion.
• Node.js, only if you want to use the MCP Inspector through npx

---

Quality Gate

Run the same checks expected before pull requests:

poetry run ruff check .
poetry run ruff format --check src tests examples
poetry run mypy
poetry run pytest -m "not integration"

On systems with make, the same gate is available as:

make quality

---

Installation

poetry install

Configuration

The server reads configuration from environment variables or a local .env file.

• MCP_DATA_ROOT: directory that contains readable datasets. Defaults to data.
• MCP_REPORTS_ROOT: directory that contains saved modeling reports. Defaults to reports.
• MCP_TRANSPORT: stdio or streamable-http. Defaults to streamable-http.
• MCP_HOST: bind host for HTTP mode. Defaults to 127.0.0.1.
• MCP_PORT: bind port for HTTP mode. Defaults to 8000.
• MCP_MAX_PREVIEW_ROWS: maximum allowed value for preview_csv. Defaults to 50.
• MCP_MAX_SQL_ROWS: reserved maximum row limit for upcoming SQL tools. Defaults to 1000.
• MCP_MAX_SQL_QUERY_LENGTH: maximum allowed SQL text length for query tools. Defaults to 20000.
• MCP_SQL_TIMEOUT_MS: maximum SQL execution time before interruption. Defaults to 5000.
• MCP_MAX_CATEGORICAL_VALUES: maximum top-value examples returned per categorical column in profile_csv. Defaults to 5.
• MCP_MAX_DATASET_BYTES: maximum readable dataset size in bytes. Defaults to 25000000.
• MCP_PROFILE_CACHE_ENABLED: enable or disable in-memory profile caching. Defaults to true.
• MCP_PROFILE_CACHE_MAX_ENTRIES: maximum cached profile entries. Defaults to 128.
• MCP_LOG_LEVEL: server log level. Defaults to INFO.
• MCP_API_KEY: optional shared bearer token for Streamable HTTP mode. Disabled by default.
• MCP_TRACING_ENABLED: enable optional OpenTelemetry spans. Defaults to false.
• MCP_TRACING_SERVICE_NAME: service name reported in traces. Defaults to ds-workspace-mcp.
• MCP_TRACING_CONSOLE_EXPORTER: print spans to stdout through the OpenTelemetry console exporter. Defaults to false.

Example:

MCP_DATA_ROOT=./data
MCP_REPORTS_ROOT=./reports
MCP_TRANSPORT=streamable-http
MCP_HOST=127.0.0.1
MCP_PORT=8000
MCP_MAX_PREVIEW_ROWS=50
MCP_MAX_SQL_ROWS=1000
MCP_MAX_SQL_QUERY_LENGTH=20000
MCP_SQL_TIMEOUT_MS=5000
MCP_MAX_CATEGORICAL_VALUES=5
MCP_MAX_DATASET_BYTES=25000000
MCP_PROFILE_CACHE_ENABLED=true
MCP_PROFILE_CACHE_MAX_ENTRIES=128
MCP_LOG_LEVEL=INFO
MCP_API_KEY=
MCP_TRACING_ENABLED=false
MCP_TRACING_SERVICE_NAME=ds-workspace-mcp
MCP_TRACING_CONSOLE_EXPORTER=false

Logging

The server emits structured logs with timestamp, level, module, and message.

• Tool and resource invocations are logged with dataset names and safe counts only.
• Validation failures are logged clearly.
• Full dataset contents and preview row payloads are intentionally not logged.

Set the log level with MCP_LOG_LEVEL, for example:

MCP_LOG_LEVEL=DEBUG

Tracing

OpenTelemetry tracing is optional production polish.

Install the optional dependencies with:

poetry install --extras opentelemetry

To enable local tracing with console output:

MCP_TRACING_ENABLED=true
MCP_TRACING_CONSOLE_EXPORTER=true
poetry run ds-workspace-mcp serve

Current spans cover dataset resolution, CSV reads, profiling, SQL query execution, and MCP tool boundaries.

If tracing is enabled without the optional dependencies installed, the server continues to run and logs a warning instead of failing.

---

Troubleshooting

• Dataset not found: ...: the file name must exist inside MCP_DATA_ROOT, and the server only sees files under that directory.
• Only CSV files are supported., Only SQLite files are supported., or Unsupported dataset format: ...: the requested file extension is not allowed for that tool.
• Access outside the configured data directory is not allowed.: the request attempted path traversal such as ../....
• Could not profile dataset: ...: the file could be opened but could not be profiled safely; validate the CSV structure and encoding.
• Could not read dataset: ...: the CSV could not be decoded or parsed safely.
• Dataset exceeds the maximum allowed size ...: the file is larger than the configured dataset-size guardrail.
• Destructive or schema-changing SQL is not allowed.: the SQL tool only accepts bounded read-only SELECT or WITH queries.
• SQL query exceeded the timeout ...: the query ran longer than the configured SQL timeout.

These errors intentionally avoid exposing absolute local paths in responses.

---

Run with Streamable HTTP

MCP_TRANSPORT=streamable-http poetry run ds-workspace-mcp

By default, the server reads datasets from ./data.

You can override that directory:

MCP_DATA_ROOT=/path/to/csv/files MCP_TRANSPORT=streamable-http poetry run ds-workspace-mcp

Then connect with the MCP Inspector:

npx -y @modelcontextprotocol/inspector

Use this URL:

http://localhost:8000/mcp

---

Run with stdio

MCP_TRANSPORT=stdio poetry run ds-workspace-mcp

This is useful when connecting the server directly to an MCP client that launches local servers.

Synthetic Dataset

The repository also includes a reproducible healthcare operations dataset generator.

Generate a fresh sample into data/:

poetry run generate-sample-healthcare-data

Custom example:

poetry run generate-sample-healthcare-data --output data/custom_clinic_usage.csv --days 180 --clinics 6 --seed 7

Generated columns:

• clinic_id
• date
• appointments_scheduled
• appointments_completed
• cancellations
• no_shows
• marketing_campaign
• local_holiday
• staff_available
• average_wait_time
• patient_satisfaction_score

The generator includes weekday effects, seasonal demand, campaign lift, holiday effects, random noise, and light missingness in a couple of operational fields.

CLI

The main console command now exposes a small local workflow CLI.

Serve the MCP server:

poetry run ds-workspace-mcp serve

poetry run ds-workspace-mcp also defaults to serve.

List available datasets:

poetry run ds-workspace-mcp list-datasets

Profile one dataset as JSON:

poetry run ds-workspace-mcp profile-dataset sample_clinic_usage.csv

Build a first-pass modeling plan as JSON:

poetry run ds-workspace-mcp plan-modeling sample_clinic_usage.csv --target-column appointments_completed

Build a markdown modeling report:

poetry run ds-workspace-mcp report-modeling sample_clinic_usage.csv --target-column appointments_completed

Save a markdown modeling report into reports/:

poetry run ds-workspace-mcp save-modeling-report sample_clinic_usage.csv --target-column appointments_completed --output-name clinic-usage-report.md

Saved reports, copied reports, renamed reports, and saved report sections fail when the target file already exists unless --overwrite is passed or the MCP tool argument overwrite is set to true.

List saved modeling reports:

poetry run ds-workspace-mcp list-modeling-reports

Search saved modeling reports:

poetry run ds-workspace-mcp search-modeling-reports clinic

Search saved modeling report content:

poetry run ds-workspace-mcp search-modeling-report-content elevated

Search the latest saved modeling report content:

poetry run ds-workspace-mcp search-latest-modeling-report-content elevated

Search saved modeling report sections:

poetry run ds-workspace-mcp search-modeling-report-sections risk

Search section headings in the latest saved modeling report:

poetry run ds-workspace-mcp search-latest-modeling-report-sections risk

Summarize recurring modeling report sections:

poetry run ds-workspace-mcp summarize-modeling-report-sections

List the most recent modeling reports:

poetry run ds-workspace-mcp list-recent-modeling-reports --limit 3

Summarize saved modeling reports:

poetry run ds-workspace-mcp summarize-modeling-reports --limit 3

Read one saved modeling report:

poetry run ds-workspace-mcp read-modeling-report clinic-usage-report.md

List section headings from one saved modeling report:

poetry run ds-workspace-mcp list-modeling-report-sections clinic-usage-report.md

List section headings from the latest saved modeling report:

poetry run ds-workspace-mcp list-latest-modeling-report-sections

Read one section from a saved modeling report:

poetry run ds-workspace-mcp read-modeling-report-section clinic-usage-report.md "Summary"

Read one section from the latest saved modeling report:

poetry run ds-workspace-mcp read-latest-modeling-report-section "Summary"

Save one section from a saved modeling report:

poetry run ds-workspace-mcp save-modeling-report-section clinic-usage-report.md "Risks"

Save one section from the latest saved modeling report:

poetry run ds-workspace-mcp save-latest-modeling-report-section "Risks"

Compare one section across two saved modeling reports:

poetry run ds-workspace-mcp compare-modeling-report-sections clinic-usage-report.md clinic-usage-v2.md "Risks"

Compare one section across the two most recent saved modeling reports:

poetry run ds-workspace-mcp compare-latest-modeling-report-sections "Risks"

Read the latest saved modeling report:

poetry run ds-workspace-mcp read-latest-modeling-report

Delete one saved modeling report:

poetry run ds-workspace-mcp delete-modeling-report clinic-usage-report.md

Rename one saved modeling report:

poetry run ds-workspace-mcp rename-modeling-report clinic-usage-report.md clinic-usage-v2.md

Rename the latest saved modeling report:

poetry run ds-workspace-mcp rename-latest-modeling-report clinic-usage-latest.md

Copy one saved modeling report:

poetry run ds-workspace-mcp copy-modeling-report clinic-usage-report.md clinic-usage-snapshot.md

Copy the latest saved modeling report:

poetry run ds-workspace-mcp copy-latest-modeling-report latest-clinic-snapshot.md

Inspect one saved modeling report:

poetry run ds-workspace-mcp inspect-modeling-report clinic-usage-report.md

Inspect the latest saved modeling report:

poetry run ds-workspace-mcp inspect-latest-modeling-report

Preview one saved modeling report:

poetry run ds-workspace-mcp preview-modeling-report clinic-usage-report.md

Compare two saved modeling reports:

poetry run ds-workspace-mcp compare-modeling-reports clinic-usage-report.md clinic-usage-v2.md

Compare the two most recent modeling reports:

poetry run ds-workspace-mcp compare-latest-modeling-reports

Preview the latest saved modeling report:

poetry run ds-workspace-mcp preview-latest-modeling-report

Generate a synthetic dataset through the main CLI:

poetry run ds-workspace-mcp generate-sample-healthcare-data --days 120 --clinics 5

Examples

The repository includes runnable MCP client examples in examples/.

Use stdio against a locally launched process:

poetry run python examples/stdio_client.py

Expected output starts with:

Resources:
{
  "resources": [

Use Streamable HTTP after starting the server:

poetry run ds-workspace-mcp serve
poetry run python examples/http_client.py

Expected output starts with:

Tools:
{
  "tools": [

The examples show how to:

• initialize an MCP client session;
• list resources or tools;
• call preview_csv, profile_csv, detect_csv_issues, and summarize_correlations.
• exit with a non-zero status and a concise stderr message if the connection fails.

---

Available MCP capabilities

Resource

datasets://catalog

Returns the list of CSV files available in the configured data directory.

databases://sqlite

Returns the list of SQLite database files available in the configured data directory.

reports://modeling

Returns the list of saved modeling report file names from the local reports/ directory.

reports://modeling/latest

Returns the most recently modified saved modeling report as markdown.

reports://modeling/{output_name}

Returns one saved modeling report as markdown for the requested report file name.

reports://modeling/latest/sections

Returns section headings from the most recently modified saved modeling report.

reports://modeling/{output_name}/sections

Returns section headings from one saved modeling report.

reports://modeling/latest/sections/{section_heading}

Returns one section from the most recently modified saved modeling report as markdown.

reports://modeling/{output_name}/sections/{section_heading}

Returns one section from a saved modeling report as markdown.

Tools

summarize_dataset

Return a compact first-pass overview of a dataset, including shape, missingness highlights, identifier-like columns, strongest correlations, and recommended next tools.

Arguments:

{
  "file_name": "sample_clinic_usage.csv"
}

aggregate_dataset

Run a bounded allowlisted aggregation over any supported tabular dataset without accepting free-form executable expressions.

Arguments:

{
  "request": {
    "file_name": "sample_clinic_usage.csv",
    "group_by": ["clinic_id"],
    "metrics": [
      {
        "operation": "mean",
        "column": "appointments_completed",
        "output_name": "avg_completed"
      },
      {
        "operation": "count",
        "output_name": "row_count"
      }
    ],
    "filters": [
      {
        "column": "appointments_scheduled",
        "operation": "gt",
        "value": 0
      }
    ],
    "order_by": [
      {
        "column": "avg_completed",
        "direction": "desc"
      }
    ],
    "limit": 10
  }
}

compare_datasets

Compare two supported datasets for structural changes and lightweight drift diagnostics. Schema differences and statistical drift are reported separately, and the response states whether the comparison used full data or a bounded head sample.

Arguments:

{
  "left_file_name": "sample_clinic_usage.csv",
  "right_file_name": "sample_clinic_usage_v2.csv"
}

The result includes added/removed columns, dtype changes, null-rate changes, row-count changes, bounded categorical cardinality changes, numeric effect-size diagnostics, categorical distribution shifts, timestamp range shifts, and the sampling policy used.

suggest_target_columns

Suggest plausible target columns for modeling, with task-type hints and reasoning.

Arguments:

{
  "file_name": "sample_clinic_usage.csv"
}

suggest_feature_columns

Suggest which columns to include, review, or exclude for a supervised modeling target.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "target_column": "appointments_completed"
}

assess_modeling_readiness

Return a compact modeling-readiness summary that combines target suggestion, feature selection, leakage review, and next-step guidance.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "target_column": "appointments_completed"
}

build_experiment_plan

Build a concrete first-pass modeling plan with starter models, validation guidance, risks, metrics, next steps, and a reproducible evaluation manifest.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "target_column": "appointments_completed"
}

build_modeling_report

Build a compact markdown report artifact for review or handoff. The structured result includes an evaluation_manifest, and the markdown includes a compact manifest section with provenance details such as dataset fingerprint, executable validation strategy, selected feature counts, runtime versions, and generation time.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "target_column": "appointments_completed"
}

save_modeling_report

Build and save a markdown report artifact inside the local reports/ directory.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "target_column": "appointments_completed",
  "output_name": "clinic-usage-report.md"
}

list_modeling_reports

Return markdown modeling reports saved inside the local reports/ directory.

search_modeling_reports

Return saved markdown modeling reports whose file names match a case-insensitive substring.

Arguments:

{
  "query": "clinic"
}

search_modeling_report_content

Return saved markdown modeling reports whose content matches a case-insensitive text query.

Arguments:

{
  "query": "elevated"
}

search_latest_modeling_report_content

Return content matches from the most recently modified saved modeling report.

Arguments:

{
  "query": "elevated"
}

search_modeling_report_sections

Return saved modeling report sections whose headings match a case-insensitive text query.

Arguments:

{
  "query": "risk"
}

search_latest_modeling_report_sections

Return section-heading matches from the most recently modified saved modeling report.

Arguments:

{
  "query": "risk"
}

summarize_modeling_report_sections

Return a compact summary of recurring section headings across saved modeling reports.

list_recent_modeling_reports

Return the most recently modified saved markdown modeling reports.

Arguments:

{
  "limit": 5
}

summarize_modeling_report_catalog

Return a compact summary of saved modeling reports, including total count, total size, and recent entries.

Arguments:

{
  "limit": 5
}

read_modeling_report

Read one saved markdown modeling report from the local reports/ directory.

Arguments:

{
  "output_name": "clinic-usage-report.md"
}

list_modeling_report_sections

List markdown section headings discovered inside one saved modeling report.

Arguments:

{
  "output_name": "clinic-usage-report.md"
}

list_latest_modeling_report_sections

List markdown section headings discovered inside the most recently modified saved modeling report.

read_modeling_report_section

Read one markdown section from a saved modeling report.

Arguments:

{
  "output_name": "clinic-usage-report.md",
  "section_heading": "Summary"
}

read_latest_modeling_report_section

Read one markdown section from the most recently modified saved modeling report.

Arguments:

{
  "section_heading": "Summary"
}

save_modeling_report_section

Save one markdown section from a saved modeling report as a new markdown artifact.

Arguments:

{
  "output_name": "clinic-usage-report.md",
  "section_heading": "Risks",
  "new_output_name": "clinic-risks.md"
}

save_latest_modeling_report_section

Save one markdown section from the most recently modified saved modeling report as a new markdown artifact.

Arguments:

{
  "section_heading": "Risks",
  "new_output_name": "latest-risks.md"
}

compare_modeling_report_sections

Return a bounded diff summary between matching sections in two saved modeling reports.

Arguments:

{
  "output_name": "clinic-usage-report.md",
  "other_output_name": "clinic-usage-v2.md",
  "section_heading": "Risks"
}

compare_latest_modeling_report_sections

Return a bounded diff summary for one section across the two most recently modified saved modeling reports.

Arguments:

{
  "section_heading": "Risks"
}

read_latest_modeling_report

Read the most recently modified saved markdown modeling report from the local reports/ directory.

delete_modeling_report

Delete one saved markdown modeling report from the local reports/ directory.

Arguments:

{
  "output_name": "clinic-usage-report.md"
}

rename_modeling_report

Rename one saved markdown modeling report inside the local reports/ directory.

Arguments:

{
  "output_name": "clinic-usage-report.md",
  "new_output_name": "clinic-usage-v2.md"
}

rename_latest_modeling_report

Rename the most recently modified saved modeling report inside the local reports/ directory.

Arguments:

{
  "new_output_name": "clinic-usage-latest.md"
}

copy_modeling_report

Copy one saved markdown modeling report into a new artifact inside the local reports/ directory.

Arguments:

{
  "output_name": "clinic-usage-report.md",
  "new_output_name": "clinic-usage-snapshot.md"
}

copy_latest_modeling_report

Copy the most recently modified saved modeling report into a new artifact inside the local reports/ directory.

Arguments:

{
  "new_output_name": "latest-clinic-snapshot.md"
}

inspect_modeling_report

Return metadata for one saved markdown modeling report.

Arguments:

{
  "output_name": "clinic-usage-report.md"
}

inspect_latest_modeling_report

Return metadata for the most recently modified saved modeling report.

preview_modeling_report

Return a bounded preview of one saved markdown modeling report.

Arguments:

{
  "output_name": "clinic-usage-report.md"
}

compare_modeling_reports

Return a bounded unified diff summary between two saved markdown modeling reports.

Arguments:

{
  "output_name": "clinic-usage-report.md",
  "other_output_name": "clinic-usage-v2.md"
}

compare_latest_modeling_reports

Return a bounded unified diff summary between the two most recently modified saved modeling reports.

preview_latest_modeling_report

Return a bounded preview of the most recently modified modeling report from the local reports/ directory.

preview_csv

Preview the first rows of a CSV file.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "rows": 5
}

profile_csv

Return row count, column count, dtypes, missing values, missing percentages, bounded column summaries for numeric, categorical, boolean, and datetime fields, and advisory data-quality diagnostics.

Arguments:

{
  "file_name": "sample_clinic_usage.csv"
}

Example response shape:

{
  "file_name": "sample_clinic_usage.csv",
  "row_count": 120,
  "column_count": 6,
  "missing_values": {"wait_time": 4},
  "numeric_columns": [
    {
      "column": "appointments_completed",
      "count": 120,
      "mean": 83.2,
      "std": 9.4,
      "min": 61.0,
      "q25": 77.0,
      "median": 84.0,
      "q75": 90.0,
      "max": 103.0,
      "iqr": 13.0,
      "robust_spread": 9.64,
      "histogram": [{"lower_bound": 61.0, "upper_bound": 70.0, "count": 4}],
      "iqr_outlier_count": 1,
      "z_score_outlier_count": 0,
      "skewness": 0.12,
      "quality_signals": []
    }
  ],
  "categorical_columns": [
    {
      "column": "clinic_id",
      "count": 120,
      "unique_count": 4,
      "top_value": "north",
      "top_value_frequency": 32,
      "top_values": [{"value": "north", "count": 32}],
      "rare_category_count": 0,
      "rare_category_mass": 0.0,
      "entropy": 1.97,
      "normalized_entropy": 0.98,
      "quality_signals": []
    }
  ],
  "boolean_columns": [
    {
      "column": "local_holiday",
      "true_count": 8,
      "false_count": 112,
      "missing_count": 0
    }
  ],
  "datetime_columns": [
    {
      "column": "date",
      "count": 120,
      "min": "2024-01-01T00:00:00",
      "max": "2024-04-29T00:00:00"
    }
  ],
  "profiling_limits": {
    "max_categorical_values": 5,
    "max_histogram_bins": 10,
    "max_candidate_key_columns": 12,
    "max_candidate_key_combinations": 25
  },
  "data_quality": {
    "duplicate_row_count": 0,
    "duplicate_row_percentage": 0.0,
    "candidate_keys": [],
    "empty_columns": [],
    "one_value_columns": [],
    "probable_free_text_columns": [],
    "possible_identifier_columns": []
  }
}

Histogram bins, top values, and candidate key searches are capped. Quality signals are advisory and include a reason plus severity and confidence metadata.

Profiling limits:

• Categorical examples are intentionally bounded by MCP_MAX_CATEGORICAL_VALUES.
• The profiler returns summaries rather than raw wide-table payloads.
• Datetime detection is conservative to avoid misclassifying free-text columns.
• Profile results are cached in memory by file path, size, modified time, and profiling options.
• The profile cache is thread-safe and reports path-free hit, miss, eviction, and entry metrics.

detect_csv_issues

Detect simple data-quality issues, such as high missingness and likely identifier columns.

Arguments:

{
  "file_name": "sample_clinic_usage.csv"
}

query_csv_with_duckdb

Run a safe read-only DuckDB query against the CSV loaded as a temporary table named dataset.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "sql": "SELECT clinic_id, AVG(appointments_completed) AS avg_completed FROM dataset GROUP BY clinic_id ORDER BY avg_completed DESC",
  "limit": 10
}

list_sqlite_databases_tool

Return the SQLite database files available in the configured data directory.

list_sqlite_tables

List user tables from a SQLite database.

Arguments:

{
  "file_name": "clinic_metrics.sqlite"
}

describe_sqlite_table

Describe the columns for a SQLite table.

Arguments:

{
  "file_name": "clinic_metrics.sqlite",
  "table_name": "visits"
}

query_sqlite

Run a safe read-only SQLite query against a database inside the configured data directory.

Arguments:

{
  "file_name": "clinic_metrics.sqlite",
  "sql": "SELECT clinic, SUM(appointments) AS total_appointments FROM visits GROUP BY clinic ORDER BY total_appointments DESC",
  "limit": 10
}

Example response shape:

{
  "file_name": "clinic_metrics.sqlite",
  "columns": ["clinic", "total_appointments"],
  "rows": [
    {"clinic": "north", "total_appointments": 22}
  ],
  "row_count": 1,
  "limit_applied": 10
}

summarize_correlations

Rank the top absolute correlations among numeric columns.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "method": "pearson"
}

detect_possible_target_leakage

Return heuristic warnings about columns that may leak the target.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "target_column": "appointments_completed"
}

validate_time_series_dataset_tool

Validate whether a dataset looks ready for forecasting or time-series modeling.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "time_column": "date",
  "target_column": "appointments_completed",
  "group_column": "clinic_id"
}

evaluate_forecast_baselines

Evaluate transparent forecasting baselines with chronological rolling-origin backtesting. Regular frequencies are required; irregular series are rejected instead of being scored.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "time_column": "date",
  "target_column": "appointments_completed",
  "group_column": "clinic_id",
  "forecast_horizon": 1,
  "test_size": 0.2,
  "seasonal_period": 7
}

The result includes last-value naive and drift baselines, plus seasonal naive when the frequency and history support a seasonal period. Metrics include MAE, RMSE, MASE when the training denominator is valid, and sMAPE with both-zero terms contributing 0. The structured result also includes an evaluation manifest with dataset fingerprint, rolling-origin validation metadata, metric definitions, and runtime bounds.

evaluate_baseline_model

Evaluate a dummy baseline model for regression, binary classification, or multiclass classification. The result includes an evaluation manifest with the dataset fingerprint, selected features, split metadata, baseline definition, metric definitions, package version, Python version, and relevant configuration bounds.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "target_column": "appointments_completed",
  "task_type": "regression",
  "test_size": 0.2,
  "random_state": 42,
  "validation_strategy": "chronological",
  "time_column": "date"
}

Prompt

dataset_analysis_prompt

Creates a reusable analysis prompt for a dataset.

Arguments:

{
  "file_name": "sample_clinic_usage.csv",
  "objective": "forecast clinic usage and staffing demand"
}

modeling_report_review_prompt

Creates a reusable review prompt for a saved modeling report.

Arguments:

{
  "output_name": "clinic-usage-report.md",
  "focus": "decide whether the report is ready for stakeholder review"
}

Use "latest" for output_name to target the most recently modified saved modeling report.

---

Test

poetry run pytest

Run static checks:

poetry run ruff check .
poetry run mypy

Development Workflow

Local development uses the same quality gate as CI:

poetry run ruff check .
poetry run mypy
poetry run pytest

GitHub Actions runs those checks on pushes to main and develop, and on pull requests, across Python 3.11 and 3.12.

Documentation Rule

README maintenance is part of the definition of done for this repository.

• Any user-facing capability change should update README.md in the same change.
• CLI, MCP tool, workflow, and setup documentation should stay aligned with the implemented repo state.
• We use the same rule as a default for other repositories unless there is a stronger repo-specific convention.

Security

For local-only work, the default configuration keeps HTTP auth disabled.

If you set MCP_API_KEY, Streamable HTTP mode requires:

Authorization: Bearer <your-key>

Limitations of this approach:

• it is a simple shared secret, not a full user or OAuth system;
• it does not apply to stdio transport;
• it should be paired with HTTPS and normal secret management if exposed beyond local development.

Packaging and Release

Build local release artifacts with:

poetry build

This produces both a wheel and a source distribution in dist/.

GitHub Actions also includes a release workflow in .github/workflows/release.yml:

• workflow_dispatch builds artifacts for a manual release dry run.
• pushing a tag such as v1.0.0 builds release artifacts from the tagged repository state.

This repository treats Git tags and GitHub releases as the primary release mechanism. Package indexes are optional and are not part of the core product workflow.

Notebooks

The repository includes two walkthrough notebooks in notebooks/:

• 01_mcp_dataset_inspection.ipynb: dataset discovery, preview, profiling, issue detection, and correlation inspection on the bundled sample CSV.
• 02_forecasting_readiness_workflow.ipynb: synthetic healthcare dataset generation, time-series validation, leakage review, and baseline regression evaluation.

To run them locally, start Jupyter from the project environment:

poetry run jupyter notebook

If Jupyter is not installed in your Poetry environment yet, install it in that environment first with your preferred tool.

Project Docs

• CHANGELOG.md: user-facing change history
• CITATION.cff: citation metadata for GitHub and Zenodo
• CONTRIBUTING.md: local development and PR expectations
• SECURITY.md: security reporting and hardening expectations
• docs/API_CONTRACT.md: versioned public MCP, CLI, and configuration contract
• docs/SECURITY_MODEL.md: trust boundaries, controls, and known security gaps
• docs/DEPLOYMENT.md: supported local, HTTP, Docker, and Compose deployment paths
• docs/ARCHITECTURE.md: component responsibilities and request flow
• docs/RELEASE_CHECKLIST.md: release prep checklist

---

Docker

Build:

docker build -t ds-workspace-mcp .

Run:

docker run --rm -p 8000:8000 ds-workspace-mcp

Mount your own datasets:

docker run --rm \
  -p 8000:8000 \
  -v "$(pwd)/data:/app/data" \
  -v "$(pwd)/reports:/app/reports" \
  ds-workspace-mcp

Run with Docker Compose:

docker compose up --build

The Compose setup reads defaults from .env.example, overrides from .env when present, binds the server to 0.0.0.0, publishes MCP_PORT, mounts local data/ at /app/data, and mounts local reports/ at /app/reports so saved reports survive container recreation.

To inspect the running HTTP server with MCP Inspector:

npx -y @modelcontextprotocol/inspector

To stop the stack:

docker compose down

---

Safety design

The server does not read arbitrary files.

All dataset paths are resolved inside MCP_DATA_ROOT, which defaults to ./data. Path traversal attempts such as ../secret.csv are rejected.

Supported tabular files are CSV, Parquet, JSON, and Excel. SQLite database files are supported through the dedicated read-only SQLite tools.

DuckDB query safety notes:

• CSV files are resolved through the existing safe data-root checks before DuckDB sees any data.
• Queries must be a single SELECT or WITH statement against the temporary dataset table.
• DuckDB external access and extension autoload/autoinstall are disabled before user SQL executes.
• Parsed SQL relation references must resolve only to the registered in-memory dataset relation.
• Destructive or schema-changing SQL is rejected.
• External file-reading functions such as read_csv(...) are rejected.
• A final row limit is always applied, with MCP_MAX_SQL_ROWS acting as the upper bound.
• SQL text is bounded by MCP_MAX_SQL_QUERY_LENGTH.
• Interruption is requested when execution or result materialization exceeds MCP_SQL_TIMEOUT_MS.

SQLite safety notes:

• Only .sqlite, .sqlite3, and .db files inside MCP_DATA_ROOT are allowed.
• SQLite connections are opened in read-only mode.
• Schema-changing and destructive SQL is rejected.
• Query execution is limited to a single SELECT or WITH statement.
• A final row limit is always applied, with MCP_MAX_SQL_ROWS acting as the upper bound.
• SQL text is bounded by MCP_MAX_SQL_QUERY_LENGTH.
• Long-running queries and result fetching are interrupted via SQLite progress handlers when they exceed MCP_SQL_TIMEOUT_MS.

Diagnostics notes:

• Correlation summaries are limited to numeric columns and return only the top ranked pairs.
• Leakage warnings include severity and confidence, and remain review signals rather than proof.
• Exact target duplicates and likely identifiers are strong exclusion evidence for baseline feature selection.
• Name overlap and strong numeric correlation are review evidence by themselves, not automatic exclusion proof.

Time-series validation notes:

• Timestamp parsing is conservative and fails if the declared time column cannot be parsed at all.
• The validator checks sorting, duplicate timestamps, inferred frequency, missing intervals, grouped gaps, and missing target values.
• Frequency inference reports a structured kind: regular, approximately_regular, irregular, insufficient_data, or grouped heterogeneous.
• Missing intervals are counted only after a regular or sufficiently supported approximate fixed interval is established.
• Grouped series infer frequency per group and avoid reporting one global frequency when groups disagree.
• History-length warnings are heuristic and intended for baseline forecasting readiness, not strict modeling requirements.
• Forecast baselines use chronological rolling-origin evaluation and require regular inferred frequencies.
• Seasonal naive runs only when an explicit or inferred seasonal period is valid for the horizon and training history.
• Forecast baseline metrics include MAE, RMSE, MASE when the denominator is valid, and sMAPE with both-zero terms contributing 0.

Baseline model notes:

• Only scikit-learn dummy baselines are used.
• regression, binary_classification, and multiclass_classification are supported.
• Validation can be random, stratified, chronological, or grouped; classification defaults to stratified validation when feasible.
• Classification baselines report full, train, and test class counts and fail clearly when a split cannot represent every class.
• Classification metrics include accuracy, balanced accuracy, macro F1, and weighted F1.
• Chronological validation requires a parseable time column, never shuffles, and reports train/test time boundaries.
• Grouped validation requires a group column and reports train/test group counts plus group-overlap status.
• These metrics are reference baselines for comparison, not final model quality targets.

---

Suggested Portfolio Positioning

An MCP server that gives AI assistants safe, structured access to local analytical datasets, SQL sources, and reusable modeling report workflows.

This project demonstrates AI engineering, data tooling, typed Python service design, MCP integration, analytical workflow ergonomics, and security-minded local tool design.