Crosstabs MCP Server

The unpublished source candidate provides two local MCP servers: 39 statistical tools for focused contingency-table analysis and 29 headless research workflow tools that take an agent from project creation and dataset import through approved survey designs, tab books, tracker repair, editable reports, and portable export. Individual statistical tools report whether their inference is exact, asymptotic, or simulated.

Both commands are local stdio servers. The package contains no HTTP/SSE transport or cloud-deployment configuration; network transport arguments fail closed. Use the separate five-tool public MCP at crosstabs.com only for bounded aggregate matrices and public evidence.

crosstabs is an aggregate calculator. Its ordinary inferential tools require caller-attested inputSemantics="integer_frequency_counts" and non-negative integer frequency cells; they do not accept respondent, calibration, raking, or complex-survey weights. Use crosstabs-headless for a completed supported saved survey design and its disclosed design-aware inference.

Inferential aggregate calls also require weighted=false. The two descriptive aggregate visual tools, stacked_bar_data and correspondence_analysis, require inputSemantics="descriptive_aggregate_values"; their fractional values are not survey-weighted inference. mosaic_plot_data remains an inferential residual diagnostic and therefore requires integer frequency counts. Missing or unsupported semantics, fractional cells, non-finite values, negative values, and unsafe integers are rejected before any statistics are returned. proportion_ci separately requires countSemantics="integer_binomial_counts" and safe integer binomial counts for successes and total.

Features

Core Statistical Tests

Effect Sizes & Measures

Ordinal Measures

Agreement & Reliability

Advanced Analysis

Installation

From PyPI (recommended)

pip install crosstabs

Inspect the public source distribution

python -m pip download --no-deps --no-binary=:all: crosstabs

The development repository is currently private. PyPI publishes the package's source archive; email support@crosstabs.com for issue reports or source-access questions.

Quick Start

Run the MCP Server

# Focused statistical calculators (Python 3.10+)
crosstabs

# End-to-end research workspace (Python 3.10+ and Node.js 22+)
crosstabs-headless

Or directly:

python -m crosstabs_mcp.server

Configure an MCP client

Add to your ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "crosstabs_workspace": {
      "command": "crosstabs-headless"
    },
    "crosstabs_statistics": {
      "command": "crosstabs"
    }
  }
}

The headless server stores project state and generated artifacts under the local application-data directory. Deterministic calculations stay on the machine. It makes no hosted-AI, account, or remote-project call. Manual open-end coding remains a browser-workspace feature rather than a local MCP operation.

Headless workflow tools

create_project, import_dataset, profile_dataset, define_row_set, define_banner, apply_filter, set_weight, define_survey_design, run_table, run_tab_book, compare_waves, undo_change, replace_dataset, detect_schema_drift, repair_schema, generate_report_pack, refresh_report_pack, export_project, list_projects, inspect_project, get_audit_history, import_project, render_project_table, update_variable_metadata, propose_transformation, review_transformation, apply_transformation, undo_transformation, and run_complex_survey_method.

Project-package import is an explicit allowed-path, checksummed full-data copy. It never overwrites a project. Rendering uses a saved table/revision/evidence ID, not an ordinary re-analysis of caller-supplied cells. Inference-changing edits retain approval, revision and replay guards. The local JSON store enforces 8 MiB per serialized project (including undo/audit) and 32 MiB per database (including replay records); parser maxima are not persistence guarantees.

Mutations use expected revisions and idempotency keys. Results include structured warnings, evidence IDs, audit records, and MCP resources for generated files. For a detected multi-select row dimension, define_row_set may additionally set multipleResponseDenominator to cases or responses; ordinary and compound row sets reject that option rather than silently applying it.

Monte Carlo workload envelope

Fixed-margin Monte Carlo accepts at most 100 rows, 100 columns, and 1,000 cells, with a total frequency of 9,007,199,254,740,991. It admits 100,000 simulations and 20,000,000 cell-simulations. It samples chunks of at most 512 simulations under a 32 MiB estimated numeric peak and a 30-second server deadline. The returned workload object records the admitted dimensions, cell-simulation product, chunk size, memory estimate, memory ceiling, and deadline. The memory estimate covers the numeric simulation path, not the MCP transport's parsed JSON body. A supplied seed is reproducible for the same table, simulation count, and package versions. MCP cancellation is observed between chunks; an individual SciPy draw cannot be interrupted mid-chunk.

Usage Examples

Once configured, Claude can perform statistical analysis:

Chi-square Test

User: Test if there's an association between treatment and outcome:
      Treatment A: 50 success, 30 failure
      Treatment B: 20 success, 40 failure

Claude: [Uses chi_square_test with matrix [[50,30],[20,40]]]
        χ² = 11.67, p = 0.0006
        Cramér's V = 0.29 (small-medium effect)
        There is a significant association between treatment and outcome.

Odds Ratio

User: Compare an adverse outcome between exposure groups:
      Exposed: 30 outcome-present, 70 outcome-absent
      Unexposed: 15 outcome-present, 85 outcome-absent

Claude: [Uses odds_ratio with matrix [[30,70],[15,85]]]
        OR = 2.43 (95% CI: 1.21-4.87)
        Exposure is associated with 143% higher odds of the outcome.

The epidemiology tools (odds_ratio, relative_risk, risk_difference, and attributable_risk) use one explicit orientation: [[exposed outcome+, exposed outcome-], [unexposed outcome+, unexposed outcome-]]. They return machine-readable zero/infinite/undefined states rather than silently continuity-correcting point estimates. Attributable/prevented fractions require causal identification assumptions; an association alone does not establish the counterfactual effect of removing an exposure.

Fisher's Exact Test

User: I have a small sample: [[3,1],[1,5]]. Is it significant?

Claude: [Uses fishers_exact with the matrix]
        p = 0.190476 (two-tailed exact)
        Not statistically significant at α=0.05.

Available Tools

Development

Run Tests

pip install -e ".[dev]"
pytest tests/ -v

Project Structure

mcp-server-python/
├── crosstabs_mcp/
│   ├── __init__.py
│   ├── headless_launcher.py # Node version check and bundled-server launcher
│   ├── headless-mcp.mjs     # Local-first 29-tool workflow MCP server
│   ├── server.py          # Main MCP server
│   └── advanced_stats.py  # Compatibility statistical kernels and contracts
├── tests/
│   ├── test_statistics.py       # Statistical behavior tests
│   └── test_reference_parity.py # Public SciPy reference parity
├── scripts/                    # Distribution verification
├── LICENSE
├── pyproject.toml
├── uv.lock
└── README.md

Requirements

• Python 3.10+
• mcp >= 1.0.0
• fastmcp >= 0.1.0
• numpy >= 1.24.0
• scipy >= 1.10.0
• pandas >= 2.0.0
• statsmodels >= 0.14.0

License

MIT License - see LICENSE for details.

Contributing

The development repository is currently private. Send corrections and proposed changes to support@crosstabs.com.

Links

• PyPI package and source archive
• Web Application
• Support
• MCP Documentation