opensharing-io/opensharing

OpenSharing is an open protocol that defines how organizations can share data and AI assets with each other without copying the underlying files. It is a Linux Foundation AI and Data project, licensed under Apache 2.0. The protocol builds on an earlier project called Delta Sharing and extends it to cover additional asset types beyond database tables. The core idea is zero-copy sharing. When one organization wants to give another organization access to a dataset, a machine learning model, or an AI agent skill, the provider's server issues short-lived access credentials scoped to just that asset. The recipient uses those credentials to read the asset directly from the provider's cloud storage (such as Amazon S3, Azure storage, or Google Cloud Storage) without the data ever passing through the sharing server itself. This avoids creating copies, which reduces storage costs and keeps data in one place for governance purposes. The protocol organizes assets into a three-level hierarchy: shares, schemas, and individual assets. A share is a named collection of assets granted to one or more recipients under a single set of credentials. A schema is a logical grouping within a share. Individual assets can be database tables in Delta Lake, Iceberg, or Parquet format, volumes, which are directories of files of any type, machine learning models with versioning and provenance information, and agent skills, which are reusable AI capabilities following the AgentSkills specification. Two more types, live callable agent services and named business entity definitions, are proposed by the community but not yet finalized. The protocol is defined as a set of REST API endpoints. Providers expose endpoints for listing shares and assets, and for issuing temporary credentials. Recipients authenticate with a bearer token and query those endpoints to discover and access what they have been granted. The Python connector and Apache Spark connector already work with OpenSharing because it is a superset of Delta Sharing, which those connectors already support. The repository contains the protocol specifications as markdown files in a spec directory. There is no application code in this repository.