Local-first data privacy tooling for two jobs: masking a real table, and generating a synthetic one that behaves like it. Around them sit schema contracts, privacy-risk scoring and drift checks. Everything runs inside your own perimeter: on your machine, in your CI, or in your browser through WebAssembly.
Triplex does not send your data anywhere. There is no server, no upload and no telemetry. Datasets, identifiers, mapping files, credentials and profiling outputs stay where you run the tool. The web UI reads your file inside the browser tab and never transmits it.
All products share one Rust engine with Arrow-compatible data contracts.
| Product | What it does | Status |
|---|---|---|
| SchemaCast | Schema inference, .triplex.yaml contracts, validation, pattern checks, format conversion |
Available |
| Anonymizer: masking | Pseudonymization with a separate mapping file: tokenization, format-preserving substitution, bucketing and partial masking per column (Tier 1) | Available |
| Anonymizer: synthetic data | New rows sampled from a Gaussian copula, a Bayesian network or independent marginals, optionally differentially private, with a fidelity and leakage report per run (Tier 2). Sketch-based synthesis from a sketch of the data as classes, which you can edit as a file (Tier 3) | Tier 2 available, Tier 3 in early form |
| DatasetCard | Privacy-risk scoring per dataset, column and record: k-anonymity, l-diversity, t-closeness and related models | Available |
| DriftGuard | Schema-regression and statistical-drift check between a baseline and a candidate, built for CI | Available |
| ModelCard | Fairness and slice metrics for a binary classifier, from a predictions file | First version available |
Faker and similar libraries invent each value on its own. That works for a name in a form field,
not for tests, demos or models that depend on how columns relate. triplex synthesize fits a
model to a real table on your machine and samples new rows that keep the correlations between
numeric columns and, with the Bayesian network, the dependencies between categories. Two limits:
it needs a real table to learn from, and it drops identifier-like columns instead of imitating
them, so mask those or fill them with a fake-value generator.
What the output is, and is not:
- Pseudonymized output is still personal data. The mapping file can reverse it.
- Synthetic output carries no privacy guarantee by default. Every run writes a report that states what was measured and what is not covered.
- The differentially private mode is implemented and tested, and is awaiting independent review. Treat its guarantee as unverified until that review is published.
- ModelCard reports are not a conformity assessment.
Install the triplex command from crates.io (needs a stable Rust toolchain),
or skip the toolchain and use the Docker image. The examples use the synthetic fixtures
in tests/fixtures/ of this repository.
cargo install triplex
triplex --version
# Or, from a checkout of this repository: build the binary into target/release/
cargo build --release -p triplex
alias triplex=target/release/triplex
# Infer a schema from a file
triplex schema tests/fixtures/sample.csv
# Validate a file against a contract (exit code 0 on a match, 2 on a breach)
triplex validate --schema tests/fixtures/sample.contract.yaml --input tests/fixtures/sample.csv
# Pseudonymize a column. The salt is read from an environment variable, never from a file.
mkdir -p out
TRIPLEX_DEMO_SALT=change-me triplex anonymize \
--recipe tests/fixtures/sample.anonymize.yaml \
--input tests/fixtures/sample.csv \
--output out/anonymized.csv --mapping-out out/mapping.csv
# Generate a synthetic copy that keeps the dependencies between columns, with a fidelity and
# leakage report in out/synthetic/
triplex synthesize \
--input tests/fixtures/sample.customers.csv \
--output out/customers.synthetic.csv \
--report-dir out/synthetic --method bayes-net --seed 7
# Compare a candidate file against a baseline
# (exit code 0 for no drift, 1 for statistical drift, 2 for a schema regression)
triplex driftguard \
--baseline tests/fixtures/driftguard.baseline.csv \
--candidate tests/fixtures/driftguard.candidate.csv \
--out-dir out/driftRun triplex without arguments for the full command list, including convert,
check-patterns, risk, modelcard and reveal-mapping. Commands that produce a report
write three files: a deterministic audit JSON, a standalone offline HTML report and a Python
script that reproduces the numbers with Polars.
docker run --rm --network none --user "$(id -u):$(id -g)" -v "$PWD:/data" \
ghcr.io/triplexanalytics/triplex schema tests/fixtures/sample.csvThe image is published for amd64 and arm64. It holds only the triplex binary and runs as a
non-root user. --network none lets Docker enforce what the tool already does: it makes no
network request. --user makes the output files belong to you, and -v "$PWD:/data" is the only
folder the container can see.
To build the image yourself: docker build -t triplex ., with
--build-arg TRIPLEX_VERSION=<version> to set what triplex --version reports.
The same engine compiled to WebAssembly, with a table view, per-column anonymization and risk analysis. The Synthesize and Sketch dialogs are part of Triplex Pro, a paid extension; synthetic data is available in the CLI and the Python package.
make -C apps/triplex-web setup # one time: wasm target, wasm-bindgen CLI, npm dependencies
make -C apps/triplex-web dev # build the wasm and start the dev serverSee apps/triplex-web/README.md for the build steps behind these targets. The UI also runs as a Jupyter widget: packages/triplex-jupyter.
The engine also runs in plain Python, with no browser and no subprocess:
pip install triplexanalyticsimport triplex
schema = triplex.infer_schema("tests/fixtures/sample.csv")
run = triplex.synthesize("tests/fixtures/sample.customers.csv", rows=1000, seed=1)Wheels cover Linux (x86_64 and aarch64), macOS and Windows, for Python 3.9 and later. See packages/triplex-python.
Guides, the glossary and the evaluation of each product are on triplexanalytics.io/docs.
| Path | Contents |
|---|---|
crates/triplex-core, crates/triplex-io |
Shared platform: data contracts and file formats |
crates/triplex-schemacast, -anonymizer, -datasetcard, -driftguard, -modelcard |
One crate per product |
apps/triplex-cli |
The triplex command-line tool |
apps/triplex-web |
WebAssembly bindings (src/) and the React UI (www/) |
apps/triplex-mcp |
Local MCP server over SchemaCast, stdio only |
packages/triplex-python |
Python bindings: import triplex |
packages/triplex-jupyter |
The web UI as a Jupyter widget |
tests/fixtures/ |
Synthetic test data only |
cargo fmt --all -- --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace
make -C apps/triplex-web test
(cd packages/triplex-python && pip install ".[dev]" && python -m pytest tests)See CONTRIBUTING.md before opening an issue or a pull request, and SECURITY.md for how to report a vulnerability.
Versions are git tags: workspace-v* for the engine and CLI, web-v* for the web UI, python-v* for the Python
package and jupyter-v* for the Jupyter widget. The
version fields in Cargo.toml, package.json and pyproject.toml are placeholders. Notable changes are listed
in CHANGELOG.md.