pytauargus-mcp

A Model Context Protocol (MCP) server for Tau-Argus, the statistical disclosure control engine. It lets an AI assistant (for example OpenCode, or any other MCP client) work on your data: you point it at a microdata file, it explores the variables, protects the tables and reports what was suppressed and why. Each tool call is one self-contained task with its full result, so a long solve can be backgrounded by the client. It is a thin wrapper over the pytauargus Python API; it re-implements no engine behaviour. Installing it also gives you the pytauargus package and the tauargus CLI.

The server runs locally on stdio: the microdata, the protection run and the results stay on the machine that hosts it, and the server makes no network calls. You can run it air-gapped, for example in a datacenter with an MCP client backed by a local LLM.

What it does

Eight tools cover the workflow.

• tauargus_version reports the pytauargus package version and the native engine version.
• tauargus_explore inspects a microdata dataset and returns its variables with their types. It reads .parquet files directly and .asc/.rda files through the engine metadata parser.
• tauargus_build_arb writes a .arb batch file from the dataset description (wraps micro_arb).
• tauargus_run_arb runs a .arb batch and reports the table files it writes.
• tauargus_protect protects one or more tables from microdata in a single call (wraps protect). The suppression method is selectable through the suppress spec (default OPT(1)); hypercube (GH) secondary suppression runs in this build for flat, non-weighted tables. It returns per-table cell counts, status counts, the unsafe and secondary cells, and all produced file paths.
• tauargus_audit audits the suppressed cells of a table computed by a batch and returns their realised feasibility intervals.
• tauargus_method_advice recommends a protection method for the table structure (dimensions, hierarchy, count response) and explains the trade-offs of the applicable alternatives. Each method is marked available when this build implements it; a method the engine cannot run is never recommended, and the result says so. This build marks global recoding, optimal, modular (HiTaS), hypercube (flat tables) and controlled rounding as available; network and CTA are not.
• tauargus_assess runs the safety rules without suppressing and returns the cells the rules flag, with their values, contributor counts and the plain-language reason each is sensitive (safety rule, frequency, singleton, zero, manual, ...), plus the applied rules in plain words. It previews what a protection run would withhold, so it runs first.

The server runs on the stdio transport.

Basis

The method advice, rule meanings and status legend describe standard statistical disclosure control practice as documented in the legacy tau-ARGUS 4.1 manual. The method trade-offs follow the overview in Meindl (2025), Statistical Disclosure Control. Improving tools and methods to facilitate the application, doctoral thesis, Johannes Kepler University Linz.

Install

Python 3.11 to 3.14, on the platforms the pytauargus wheel covers (macOS arm64, Linux x86_64 and aarch64, Windows x86_64). The wheel bundles the native engine and HiGHS, so no compiler, CMake, R or Java is needed.

pip install pytauargus-mcp
# or, as an extra of the engine package:
pip install "pytauargus[mcp]"

To try it without installing anything (requires uv), run it with uvx:

uvx pytauargus-mcp

Run

tauargus-mcp   # starts the server on stdio

The server is a local process that the client starts for you. Point the client at the tauargus-mcp console script. With a pip install it is on the PATH; without one, use uvx pytauargus-mcp instead.

OpenCode. Add a tauargus server to the project or user config:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "servers": {
      "tauargus": {
        "type": "local",
        "command": ["tauargus-mcp"]
      }
    }
  }
}

Claude Desktop. Add a tauargus entry to claude_desktop_config.json (the app opens it from Settings, Developer, Edit Config) and restart it:

{
  "mcpServers": {
    "tauargus": { "command": "tauargus-mcp" }
  }
}

Desktop starts the command without your shell environment; if it does not find the script, use its absolute path (which tauargus-mcp).

Claude Code.

claude mcp add tauargus -- tauargus-mcp

Gemini CLI.

gemini mcp add -s user tauargus tauargus-mcp

or add the same mcpServers entry to ~/.gemini/settings.json.

ChatGPT. Its custom connectors connect to a remote server over HTTPS and cannot start a local one, so this server does not work with ChatGPT.

When the server connects, the client lists its eight tools. In OpenCode, /mcps shows ✓ tauargus connected.

Develop (monorepo)

cd mcp
uv sync --extra test
.venv/bin/python -m pytest tests -q

The uv sync step installs the parent package editable and builds its native extension, so the pinned engine checkout's superbuild output (engine/build/lib) must exist first. scripts/sync-venv.sh builds it if needed (it reads the sibling ../tauargus-engine checkout; point TAUARGUS_ENGINE_DIR at it for the mcp venv too, so the editable build uses the same checkout). The tests read their fixtures from $TAUARGUS_ENGINE_DIR/data (default: the sibling checkout).