Register in a project

Save the example function in a project to use it through the admin UI and AI. Python, CLI commands and the server read the same project files.

Create a project

Run in the environment where you installed Decision Layer.

decision-layer init my_team
cd my_team
python -m pip install -e '.[test]'

Two Skills and two Python Methods are created as editable project files:

my_team/
├── decision-layer.yaml
├── pyproject.toml
└── my_team/methods/
    ├── metric_lookup/SKILL.md
    ├── drilldown/SKILL.md
    ├── ols/
    │   ├── method.py
    │   └── test_method.py
    └── matching/
        ├── method.py
        └── test_method.py

• SKILL.md contains instructions for AI queries and analysis.
• method.py contains the calculation and METHOD = python_method(...) registration.
• test_method.py checks known synthetic answers and invalid inputs.
• pyproject.toml declares project dependencies.

OLS uses statsmodels; matching uses scikit-learn. Results describe observed associations and group differences, without establishing causality.

Check and serve

decision-layer check
decision-layer test
decision-layer list
decision-layer serve --reload

Command	Purpose
check	Validate registration and references, without querying or calculating.
test	Run 13 OLS and matching tests, without Cube or a server.
list	Show names, versions and source paths.
serve	Start the admin UI and API with the same project.

The default names are analysis.metric_lookup, analysis.drilldown, stats.regression_adjustment and stats.binary_matching. New projects discover these local files. Remove a definition folder if it is not needed.

Open the printed admin URL, normally http://127.0.0.1:8000/ui/. --reload restarts the server when files change. Run the tests and increment the Method version when changing a calculation. From another directory, use decision-layer serve --project /path/to/my_team.

Add a Method

decision-layer new python rank
decision-layer new skill metric_review

These create my_team/methods/rank/method.py and my_team/methods/metric_review/SKILL.md. Use a Recipe to reuse a sequence of existing Methods. Start without defaults with decision-layer init my_team --empty.

Test in the admin UI

Connect Cube, then:

1. Select your registered Method under Methods.
2. Choose metrics, dimensions and dates under Choose inputs and test.
3. Inspect results and validation, then open Runs for queries and input evidence.

Notebook example data and local results are not uploaded automatically. The server queries the configured Cube source.

Use through AI

Follow MCP setup to connect to the same API. The default server uses:

DL_API_URL=http://127.0.0.1:8000
DL_WEB_URL=http://127.0.0.1:8000/ui

AI inspects registered Methods and input contracts, then selects data from the semantic catalog. The server records execution inputs, results and queries in a Run.

Share with your team

Keep project files in your team's Git repository. Deploy reviewed files and Python dependencies to the server, set DL_PROJECT_DIR and restart to load the project.

Runs are stored in .decision-layer/decision_layer.db and Recipes in recipes/. Notebook, CLI and server share the loader but retain their own execution history. Web-published versions are managed separately and do not overwrite project source.

Package upgrades never overwrite project files or add defaults to ordinary existing projects. An older empty managed demo gains starter files when demo runs again. Use a new directory to repeat first-install QA. Older configurations retain their built-in catalog; new configurations set include_builtin_methods: false to discover local project definitions instead.