How It Works
KitOps packages models, datasets, code, agent skills, and configuration files into versioned, signable OCI artefacts called ModelKits. These artefacts are stored in any container registry, such as Docker Hub, ECR, GCR, Harbor, Artifactory, or Jozu Hub, ensuring compatibility with existing infrastructure. ModelKits are built using the Kit CLI, a command-line tool that mirrors familiar Docker commands like kit pack, push, pull, and tag, reducing the learning curve for teams already familiar with container workflows.
Each ModelKit is a single, immutable OCI artefact that bundles all components of an AI project—models, datasets, agent skills, and policies—into a tamper-evident package. The tool generates SHA-256 hashes for every component, ensuring cryptographic integrity. Signatures are applied using Cosign, the same tool used for securing container images, and ModelKits can generate SBOM-ready evidence for compliance tracking. This approach extends the security and governance practices already in place for software containers to AI artefacts.
ModelKits support selective extraction of components, allowing teams to pull only the assets they need for testing, integration, or deployment. The packaging process is automated through GitHub Actions or other CI/CD tools, enabling teams to build and push ModelKits on every commit. Access controls, scanning, and compliance workflows from existing registries apply seamlessly to ModelKits, maintaining consistency across the AI supply chain.
KitOps also integrates with tools like MLflow and Jozu Hub for vulnerability scanning and policy enforcement. The AI Bill of Materials (AI BoM) generated by KitOps maps directly to regulatory requirements such as the EU AI Act and NIST AI Risk Management Framework, ensuring compliance without proprietary lock-in. The tool’s design aligns with the CNCF ModelPack specification, making it a standards-based solution for packaging and versioning AI projects.
Use Cases
KitOps addresses the challenges of managing AI projects where models, datasets, and configurations are often scattered across repositories, zip files, or Git submodules. By consolidating these components into a single ModelKit, teams eliminate the need for manual coordination and reduce the risk of version mismatches or tampering. For example, a data science team can package a model, its associated datasets, and configuration files into a ModelKit, then hand it off to application developers or DevOps engineers without ambiguity about which versions are being used.
In enterprise environments, KitOps ensures that AI artefacts—such as agent skills, MCP servers, and guardrail configurations—are versioned and secured alongside container images. ModelKits can be stored in private registries, ensuring that sensitive AI assets remain under organizational control while still benefiting from the same governance as software artefacts.
The tool’s ability to generate AI Bills of Materials and cryptographic proofs of integrity aligns with regulatory requirements, reducing the burden of manual compliance checks. Additionally, ModelKits can be used to automate handoffs between teams, such as moving a model from development to production, with full confidence in the artefact’s integrity.
For organizations using MLOps tools, KitOps provides a standards-based alternative to proprietary formats. Many MLOps platforms rely on closed ecosystems that can create lock-in risks or compatibility issues when tools are replaced or acquired. By using the OCI standard, ModelKits ensure interoperability across tools and teams, making it easier to collaborate regardless of the specific technologies in use.
Key Features & Capabilities
- Standardized Packaging: ModelKits bundle all AI project components—models, datasets, code, and configurations—into a single OCI artefact. This standardizes how AI assets are stored, versioned, and distributed, mirroring the workflows already established for container images.
- Cryptographic Security: Every component in a ModelKit is hashed with SHA-256, and the entire artefact is signed using Cosign. Signatures and hashes can be validated independently, providing a security posture equivalent to that of container images.
- Registry Agnosticism: ModelKits can be stored in any OCI-compliant registry, including Docker Hub, AWS ECR, Google GCR, Harbor, Artifactory, and Jozu Hub. This avoids vendor lock-in and allows teams to use their existing infrastructure for storage, access control, and scanning.
- Automated Compliance: KitOps generates AI Bills of Materials that document all components in a ModelKit, including dependencies and configurations. These BoMs can be mapped to regulatory frameworks like the EU AI Act and NIST AI RMF, simplifying compliance reporting and audits.
- Integration with Existing Tools: The Kit CLI commands (pack, push, pull, tag) are designed to be intuitive for teams familiar with Docker. Additionally, KitOps integrates with Python via PyKitOps, allowing seamless incorporation into MLflow and other data science pipelines. CI/CD automation is supported through GitHub Actions and other major tools.
- CNCF ModelPack Compliance: KitOps implements the CNCF ModelPack specification, ensuring alignment with an open standard for AI packaging. The tool can package projects as ModelPacks using the kit pack --use-model-pack command, maintaining compatibility with the broader ecosystem while offering additional enterprise features.
- Enterprise Support: Organizations requiring dedicated support for ModelKits and the Kit CLI can obtain enterprise-grade assistance from Jozu, ensuring scalability and reliability in production environments.
- Efficiency for Large Files: Unlike Git, which struggles with large files (e.g., models and datasets exceeding hundreds of GB), KitOps treats these artefacts as first-class citizens. ModelKits avoid the inefficiencies and risks associated with Git LFS, where clones may not match the original due to external dependencies on LFS servers.
- Complement to Containers: While ModelKits are not replacements for Docker containers, they can generate containers automatically when needed. This flexibility allows teams to deploy models in containers or as standalone artefacts, depending on the use case. For example, datasets may be more efficiently accessed outside containers for training or validation.
- Quick Start and Adoption: Getting started with KitOps involves packaging a model, notebook, and datasets into a ModelKit, sharing it via a registry, and selectively extracting components as needed. The tool’s design ensures minimal disruption to existing workflows, with documentation and community support available to accelerate adoption.