Open Horizons Backstage AI Catalog

Backstage AI catalog

Model AI context as governed catalog entities while keeping skill and rule content in its source repository.

When to invoke

• "Catalog our AI skills and governance rules in Backstage."
• "Register an MCP server as an API entity."
• "Add the AiResource kind to Backstage."
• "Model ownership and dependencies between AI resources."

Procedure

1. Confirm the target Backstage version and catalog backend mode.
2. Verify the current AI catalog documentation and package version.
3. Install @backstage/plugin-catalog-backend-module-ai-model in packages/backend and register the module through the backend system.
4. Model a skill or rule as kind: AiResource with lifecycle, owner, system, and type-specific fields.
5. Point backstage.io/source-location at the source skill or rule; do not copy its complete content into the entity.
6. Model an MCP server as kind: API, spec.type: mcp-server, with one or more explicit remotes.
7. Validate entity schemas, references, ownership, lifecycle, URLs, and catalog ingestion.
8. Protect internal or non-production MCP endpoints with repository policy and network controls.

Model boundaries

• AiResource type skill may declare disciplines, categories, agents, and dependsOn.
• AiResource type rule requires a category and rationale.
• mcp-server is an API subtype and uses remotes, not an embedded OpenAPI definition.
• Source content remains external and addressable through source-location annotations.

Open Horizons integration

• Scope the catalog output to the Developer IDP or Agent IDP objective and current Horizon stage.
• Preserve Open Horizons Backstage ownership, managed-identity, AKS, and evidence boundaries where applicable.
• Route cross-domain sequencing through open-horizons-orchestration (skill).

Output template

## Backstage AI catalog result

| Entity | Type | Owner | Source or remote | Validation |
| --- | --- | --- | --- | --- |

### Module
- Package:
- Backend registration:

Quality gate

• The AI catalog module and target version are verified.
• Every AI resource has stable identity, lifecycle, owner, and source location.
• Skill dependencies and rule rationales are explicit.
• MCP remotes use valid transport types and URLs.
• Catalog ingestion validates the registered model.
• Sensitive internal endpoints and credentials are not exposed.

References

• AI in the Software Catalog