Build a package
A package is how Curio ships nodes: a folder named <packageId>@<major>, such as curio.weather@1, whose manifest declares one or more node templates and the libraries they need. Every node comes from one, including the built-ins in curio.builtin@1. Packages travel as .curio.zip archives, and an install's shared catalog offers them to all of its users.
What a package contains
manifest.json: the id, name, publisher, version and compatibility, the dependencies, and one entry per node template with its label, icon, category, input and output port types, editor, and the behavior key that implements it. Its schema isdocs/schemas/node-package.v4.json.sources/: one starter file per template and the package's Python modules, or the TypeScript hooks of a custom-UI node.scripts/: the compiledbehaviors.jsthat a custom-UI node needs, named by the manifest'sbehaviorScript.skills/: what the package tells Curio's agents about its nodes, one folder per skill (see Skills for the agents).models/: in a model package, the trained model's weights (see Model Catalog).README.md, shown in the catalog, andLICENSE.integrity.json: a SHA-256 hash of every other file, written on install. After editing a package in place, regenerate it (see Write a node).
A node's logic is in its package's functions, and the node's code holds its settings and a short call of one. A module of the package offers the functions its __all__ names, or else its public functions, and each says what it does in the first paragraph of its docstring (Write a node shows how to write it). The header above each node's code lists them under Functions, so a package's nodes list their own package's functions, and the package's View details lists them under Modules. An agent that writes a node of a template reads them, and those of the packages it needs, with the template's description and its starter.
Skills for the agents
A package can tell Curio's agents how to use its nodes in skills: the folders of its skills/, one per skill, each holding a SKILL.md in the Agent Skills format with any files it reads beside it (Add skills and agents shows the format). Every agent on a dataflow that uses the package reads them: its chat lists them after its own skills and loads one when a task matches its description. A run that writes one of the package's nodes is handed, from its first message, each skill whose metadata names that node's template in curio-templates, by the template's id in the manifest:
metadata:
curio-templates: raster-calculator rasterize-buildingsA skill names only its own package's templates, and agents read it as its author wrote it. A skill can also name, in curio-asks, words of a chat message that hand it to the chat (Write a skill). Name your skills after your package: a skill named like one of Curio's, or like a skill of a package the dataflow lists before yours, is left out. Rasterization's skill, rasterization, and the Vega-Lite package's, vega-lite-charts, are examples.
Icons, tags and data types
A template's iconRef names its icon in the palette: any icon of Font Awesome's free solid or brands set, by the name fontawesome.com gives it, such as fa-solid:tree or fa-brands:github. A name Font Awesome does not have shows a cube.
The projects page files each dataflow under tags and data types. Curio finds most of them in the dataflow itself, such as Pandas for code that imports it. For what your nodes bring that their code does not show, declare dataflowCategories, at the top of the manifest for every node of the package, or in a template for its nodes:
"dataflowCategories": { "tags": ["Hydrology"], "dataTypes": ["Rasters"] }A dataflow with one of those nodes is then under Hydrology and Rasters. A name the rail already lists, such as Rasters, files the dataflow there, and a new name, such as Hydrology, adds an entry. A model package declares them the same way.
Dependencies
dependencies.python maps each library to a version spec: a PEP 440 range such as >=2.0, a bare version for an exact match, a caret such as ^0.14 for compatible releases, or an empty string for any version. Curio's own requirements cover only the framework, so even curio.builtin@1 declares its data libraries here.
A package saved from the canvas takes its dependencies from the imports in its source, mapping common names to what they install as (sklearn to scikit-learn, cv2 to opencv-python). An imported archive keeps what its author declared. Adding a package installs the libraries that are not already installed; on a deployment, each signed-in user's libraries go to their own environment.
The .curio.zip archive
The download icon on a package's row in the Node Catalog dropdown saves <packageId>@<major>.curio.zip. Export in a package's details does the same, even for a catalog package you never added, such as one you are writing under packages/. The archive holds the package's files, such as the manifest, sources/, skills/ and scripts/, at its root with no wrapper folder. integrity.json is left out and rebuilt on install, and a custom-UI package carries its compiled bundle, so the recipient needs no build step.
Import package, in the Node Catalog drawer's footer and on the Node Catalog page, installs an archive. The installer validates the manifest and each skill, rejects files outside the folders a package may contain, and refuses a package you already have: remove that one first. Install packages you trust, as with any other dependency.
Publishing, versions and forks
Publish, on a package's row in the Node Catalog dropdown or in its panel on the Node Catalog page, copies it from your store into the install's shared catalog, the packages/ folder, where every user of that install can add it. Only the account that published a package can Unpublish it. Both are hidden when Curio runs with --no-allow-publish, as the Docker deployment does; a deployment's catalog is the packages/ folder of the commit its image was built from. There is no hosted package registry.
Bump version for fixes and additions. For a breaking change, such as a renamed behavior key or a changed port type, bump compatibility.major and the folder suffix; two majors install side by side. A saved dataflow names each node as <packageId>/<templateId>, which resolves to the highest installed major, and a trailing @<major> pins one.
A fork names its parent in the manifest's lineage.forkedFrom and the original in lineage.root, and the Node Catalog dropdown groups forks under their original. A package marked "readOnly": true, such as curio.builtin@1, cannot be saved into: Save as package node copies its node into a new package of your own, which keeps neither a custom interface nor the package's skills.

