Licensing¶

modeling-api uses a split licence. The copyleft licence does not apply to persons who write models and integrations. The copyleft licence protects the hosted platform.

Component	Where	Licence
SDK and contracts: `mapi-spec`, `mapi-sdk`, `mapi-pymc`, `mapi-sklearn`, `mapi-jax`	`packages/spec`, `packages/sdk`, `packages/sdk-pymc`, `packages/sdk-sklearn`, `packages/sdk-jax`	Apache-2.0
Built-ins: steps, plots, model plugins, workflow runtime, conformance corpus	`packages/steps-core`, `packages/plots-core`, `plugins/*`, `packages/workflows`, `conformance/`	Apache-2.0
Tools: CLI, Python client, GCS emulator	`packages/cli`, `packages/client`, `packages/gcs-shim`	Apache-2.0
Platform: API server, run harness, web apps, images, infrastructure	`packages/server`, `packages/runner`, `apps/*`, `images/`, `infra/`, `ops/`, `security/`	AGPLv3 or a commercial licence
Documentation	`docs/`	CC BY 4.0

Each package gives its SPDX expression in its pyproject.toml (Apache-2.0, or AGPL-3.0-only OR LicenseRef-mapi-commercial). Each package also has its own LICENSE file. The full texts are in the repository root: LICENSE.md, LICENSE-APACHE.txt, LICENSE-AGPL.txt and COMMERCIAL-LICENSE.md.

What this means in practice¶

Plugins you write are yours

A step, model or plot recipe that you write with mapi-sdk, mapi-pymc, mapi-sklearn or mapi-jax uses Apache-2.0 libraries. You can use the licence that you want for it, and a proprietary licence is also possible.

Applications that call the platform are not derivative works

Your application can call the platform over HTTP, with the Python or TypeScript client, through webhooks or through MCP. These calls do not make your application subject to the AGPL.

If you modify and host the platform

If you modify the platform components (server, runner, apps) and offer them as a network service, the AGPLv3 says that you must publish your modifications. This does not apply if you hold a commercial licence. For commercial terms, send an email to mattreda@mattreda.pro.

Third-party libraries¶

The SDK uses xarray, numpy, pandas, zarr and Pydantic. The model packages use PyMC, ArviZ, scikit-learn, xgboost and skops. The marketing-mix models run on mmm-framework (Apache-2.0). mmm-framework adds NumPyro and JAX (Apache-2.0) and nutpie (MIT). The time-series models (ts.sarimax, ts.var, ts.ets and the other time-series models) run on statsmodels (BSD-3-Clause). The plot recipes use Altair and Plotly. From v0.13, ts.bsts, the causal models (mapi-plugin-causal, Apache-2.0) and the Pathfinder engine use pymc-extras (Apache-2.0). From v0.14, the runner decodes images with Pillow (MIT-CMU), and the classifiers of the gbm extra use LightGBM (MIT) and CatBoost (Apache-2.0). From v0.15, the deep learning models (mapi-jax) use JAX, Flax and Optax (Apache-2.0). The base-dl image also has the NVIDIA CUDA runtime wheels from pip, which have the NVIDIA licence (NVIDIA Software License Agreement). THIRD-PARTY-NOTICES.md lists them.

The visual workflow builder in the web app uses @xyflow/react (MIT). The API server connects to PostgreSQL through psycopg (LGPL-3.0). The server uses psycopg without modification, as a library that you install separately. Each library keeps its own licence.

THIRD-PARTY-NOTICES.md in the repository root is the full list. It gives each Python and JavaScript package in the container images, with its version and SPDX licence. It also contains the NOTICE files of Apache-2.0 packages. It also contains the licence texts of the JavaScript packages that the build compiles into the web app and the plot sandbox.

A script makes this file from uv.lock and pnpm-lock.yaml. After you change a lock file, run make notices. If the file is not current, make release-check and the test suite fail. Each image contains this file and the licence files above, in /usr/share/doc/mapi/.

Contributions¶

The project accepts contributions under the Contributor Licence Agreement (CONTRIBUTOR-LICENSE.md). With this agreement, the project can continue to offer the platform under the two licences. Refer to Contributing.