How to build multi-module project with Poetry?

[siddhsql]

I understand Poetry can be used for dependency management of a single project. But what if I am developing a main application that depends on other first-party libraries that are also owned by me? To build the main application I need to first build the first-party dependencies in a specific order. In Java, Maven does this for me. Is there anything like that in Python?

[Mike-Animal-Counseling]

For ordinary Python libraries, give each library its own pyproject.toml and declare its direct dependencies. Poetry can then install the whole graph from the application.

For example, with sibling directories app/, my-api/, and my-core/:

# app/pyproject.toml — alongside the rest of its package metadata
[tool.poetry.dependencies]
python = "^3.12"
my-api = { path = "../my-api", develop = true }

# my-api/pyproject.toml — a separate file
[tool.poetry.dependencies]
python = "^3.12"
my-core = { path = "../my-core", develop = true }

Running poetry install in app/ installs both libraries too. develop = true makes those local installs editable, so changes to their Python sources are visible without rebuilding wheels first.

For the Maven comparison, the distinction is what you mean by “build”: poetry build creates a distribution for the current project. It does not walk this graph and produce release artifacts for every project. If another module must generate code or compile native artifacts before your app can build, put that ordering in a build script/CI job, or look at a build orchestrator such as Pants.

One release caveat: with these path-only declarations, the app's wheel contained a file:// requirement pointing into my checkout. That path won't exist on another machine. For portable release artifacts, make sure the wheel references versioned dependencies available from your package index.

I checked an app → api → core example with Poetry 2.4.3: installing from the app made all three importable; building there produced only the app's wheel, whose metadata contained a local path to the API library. That test used pure Python packages, not custom/native build steps.

References: path dependencies and poetry build.

AI-assisted; the three-package install, imports, build and wheel metadata were checked locally.