Skip to main content
A Repository is a shared container for versioned AI assets in TrueFoundry. It is similar to a Git repository, except it stores AI and ML data instead of source code. Repositories are a common concept across AI Engineering and AI Gateway. The same repository can hold assets used by training jobs, model deployments, prompt management, and agent workflows.
Some older SDK fields, manifests, and pages use the term ML Repo or ml_repo. These refer to the same Repository concept.
Diagram showing TrueFoundry repositories as containers for versioned AI and ML assets

​
What repositories can hold

Repositories store versioned assets and the metadata needed to discover, audit, and reuse them.

Prompts

Saved prompt templates, prompt versions, input variables, model configuration, guardrails, and structured output settings.

Skills

Versioned agent skills, including SKILL.md instructions and optional supporting files such as references, scripts, and assets.

Artifacts

Versioned files and directories such as datasets, checkpoints, evaluation outputs, and other project artifacts.

Models

Versioned model files, model metadata, metrics, schemas, and lineage from training or fine-tuning workflows.

Runs

Experiment runs with parameters, metrics, tags, and links to artifacts, models, and prompts created or used during the run.
Diagram showing repository hierarchy with runs, models, artifacts, prompts, parameters, metrics, and tags

​
Runs and experiment metadata

In AI Engineering workflows, a run represents one experiment, such as training a specific model with a fixed set of hyperparameters. Runs can log:
  • Parameters — hyperparameters or configuration values, such as learning_rate or cache_size.
  • Metrics — evaluation values, such as accuracy, f1_score, or loss.
  • Tags — labels for filtering and grouping, such as env: development.
Runs can also link to models, artifacts, and prompts that were created or used during the experiment.

​
Blob storage backing

Every repository is backed by blob storage such as Amazon S3, Google Cloud Storage, Azure Blob Storage, MinIO, or another S3-compatible store. The repository metadata lives in TrueFoundry, while asset contents are stored in the configured blob storage. This means:
  • You need at least one blob storage integration before creating a repository.
  • A repository points to one storage backing and path.
  • Multiple repositories can use different paths in the same blob storage.
  • Repository assets remain in storage controlled by your organization.
For a setup walkthrough, see Create a repository.

​
Access control

Repository access controls who can view, create, update, and use the assets inside the repository. Prompts, skills, artifacts, and models do not need separate access policies for normal use; they inherit access from their parent repository. Use repository access to answer questions like:
  • Who can discover and use a prompt or skill?
  • Who can publish a new prompt, skill, artifact, or model version?
  • Who can view model files, artifact contents, and other stored data?
  • Who can manage repository settings or delete repository data?
At a high level:
Permission areaWhat it controls
Read repository/dataView repository details and read assets such as prompts, skills, models, and artifacts.
Write dataPublish or update assets by creating new versions.
Manage repositoryUpdate repository configuration and manage access.
Delete data/repositoryDelete assets or the repository itself, depending on the granted role.
For the complete permission list, see the Repository permissions section in roles and permissions.

​
When to create separate repositories

Create separate repositories when assets need different ownership, access control, storage locations, or lifecycle policies. Common patterns include:
  • By team — for example, support-ai, growth-ai, or ml-platform.
  • By environment — for example, checkout-dev and checkout-prod.
  • By sensitivity — for example, keeping regulated model artifacts or sensitive prompts in a restricted repository.
  • By project — for example, grouping prompts, skills, artifacts, and models for a single agent or application.