Agent Paranoid Android

Safety-first, deterministic synthetic test data generation from CSV structure, safe profiles, reviewed DatasetSpec files, and allowlisted Trino metadata. The CLI and Python library are primary; MCP, Trino, and AI providers are integrations. Source rows are profiled, never shuffled or copied into generated output.

Read the documentation for tutorials, concepts, configuration, MCP setup, and troubleshooting.

Current version: 1.0.0rc2. Package: agent-paranoid-android; CLI: test-data-agent.

What It Preserves

From bounded evidence and a reviewed DatasetSpec, generation can preserve schema and types, nullability, ranked distribution and scale shape, approved FK graphs, temporal dependencies, and executable business rules. AI may propose relationships and rules; human review and deterministic validation remain the authority boundary.

It intentionally does not preserve or copy source values or real PII. It does not certify statistical anonymity, protection from every re-identification attack, or cross-environment byte identity. A seed provides logical reproducibility under the recorded package, dependency, locale, and serializer environment.

Install

Python 3.11 or newer is required. CI tests CPython 3.11 through 3.14.

python3 -m pip install agent-paranoid-android
test-data-agent doctor

Install only features you use:

python3 -m pip install "agent-paranoid-android[parquet]"
python3 -m pip install "agent-paranoid-android[mcp,trino]"
python3 -m pip install "agent-paranoid-android[openai]"

First Offline Run

Run the installed package with its bundled fictional customer fixture. No checkout, network, Trino, MCP, or provider is required:

test-data-agent demo --output out/demo

A successful run reports:

Generated synthetic dataset: out/demo | rows: customers=12 | seed: 20260801 | validation: passed | source rows copied: no

Representative deterministic output:

customer_id,email,segment,signup_date
syn_customers_00000001,amber21@example.test,category_1,2024-02-10

The demo preserves evidenced column names and types, non-null shape, category rank, and date range. Its fixture has no relationship or business-rule evidence, so the demo makes no claim about those properties. The output folder contains:

• customers.csv;
• csv_profile.json;
• dataset_spec.json;
• validation_report.json;
• generation_manifest.json.

The destination must not already exist. Review the manifest and effective spec before accepting any dataset. Then follow First CSV Dataset to profile your own input.

Choose A Guide

Safety

The project derives bounded metadata such as field types, null ratios, ranges, masked patterns, and safe low-cardinality distributions. It rejects or bounds:

• raw detected PII, credentials, tokens, and private keys in profiles;
• source-row copying and source/output path reuse;
• path traversal and symlink escapes through generator MCP tools;
• unrestricted SQL and write operations through Trino tools;
• oversized input, output, rule, query, and generation work.

Human review is still required for ambiguous identifiers, rare free text, inferred relationships, and organization-specific privacy policy.

Development

python3 -m pip install "uv==0.11.23"
uv sync --frozen --all-extras --no-install-project
uv sync --frozen --all-extras --no-editable --no-build-isolation
uv run --no-sync scripts/check_release.sh

See Contributing, Support, Governance, Code Of Conduct, Security Policy, Changelog, and License.

Releases use tokenless PyPI Trusted Publishing with verified wheels/source distributions, checksums, SBOMs, and GitHub attestations.

AI-Assisted Development

AI-assisted changes require human review and tests; never send production data, raw PII, credentials, or tokens to AI. The name nods to Radiohead's "Paranoid Android"; this project is unaffiliated.