port-labs Build a context lake
A context lake is Port's data model once it stops being a flat catalog and
becomes a connected graph: blueprints related to each other, with mirror and
aggregation properties, descriptions, and relation titles precise enough
that an AI agent can traverse it and act, not just browse it. This skill
designs that model. For the JSON mechanics of any single blueprint,
property, or relation, use port-blueprints; this skill is the layer above
that: what to build, in what order, and why.
Prerequisites
- Go over the
port-getting-startedskill first if this is your first time working with Port, it covers signing up and connecting Port's MCP server. - If Port's MCP server is connected, this skill can use it to check what
already exists (
list_blueprints) and apply the resulting blueprint, relation, and property changes directly (upsert_blueprint), rather than just handing you JSON to apply yourself. Searchsearch_port_knowledge_sourcesfor anything this skill doesn't cover.
How to build it
- Start from what Port already gives you. Every account ships built-in
blueprints (
_user,_team,_scorecard, and others) that you extend, not replace. See references/default-blueprints.md. - Design the core layer before the provider layer. Model your organization's own shape first (organization, team, service), then relate it to provider-specific blueprints your integrations create (GitHub, GitLab, Jira, ...), rather than duplicating provider data into your core blueprints. See references/reference-architecture.md for a full worked example.
- Model hierarchy with a self-relation, not one blueprint per tier. A
parent_teamrelation on_teamhandles arbitrary org depth. See references/relations-and-hierarchy.md. - Use ownership instead of a custom team relation. Port's built-in
$teamfield already rolls up into_teamaggregations for free. See references/ownership.md. - Add mirror and aggregation properties, not duplicated data. Mirror a
single related value onto its source; aggregate across many related
entities on the blueprint that sits above them. For the JSON shape, see
port-blueprints' references/calculation-properties.md; for which one to use and where to put it, see references/mirror-vs-aggregation.md. - Write it as an ontology, not a schema. Descriptions, relation titles, and typed properties are what let an agent understand the graph instead of just enumerating it. See references/ontology.md.
- Decide integration vs. MCP connector per data source. Only data that needs relations, ownership, or scorecards belongs in the lake as ingested entities; everything else can stay a live MCP fetch. See references/ingestion-strategy.md.
Examples
- assets/team-hierarchy-patch.json: adds
a
parent_teamself-relation and mirror property to_team. - assets/service-github-mirrors-patch.json:
relates
servicetogithubRepositoryand mirrors its key fields. - assets/team-aggregation-patch.json:
adds an aggregation property rolling up owned services onto
_team.
Common pitfalls
| Symptom | Cause | Fix |
|---|---|---|
Tried to delete or fully redefine _user/_team |
System blueprints can't be deleted, only extended | PATCH in new relations only; leave the protected schema alone |
service has no team relation and it seems wrong |
Port uses built-in ownership ($team), not a custom relation, for this |
Set ownership on service, don't add a redundant relation |
| One blueprint per org-hierarchy tier (team, group, division, org) | Hierarchy modeled as a chain of distinct blueprints instead of a self-relation | Use a single parent_team self-relation on _team, extensible to any depth |
| Same repo/PR data duplicated on two different blueprints | Two integrations model the same real-world thing separately | Map by property into the existing blueprint instead of creating a second one |
| Agent can't tell which of several relations to a blueprint is which | Relations exist but have generic titles like team or service |
Give each relation a semantic title (Owned by, Runs in, Depends on) |
Quick reference
- Core layer first (organization, team, service), provider layer relates to it, never duplicates it.
- Hierarchy: one self-relation (
parent_team), not one blueprint per tier. - Ownership over relations: use
$team, don't hand-roll a team relation. - Mirror = one related value, surfaced as-is. Aggregation = computed across many, placed on the higher-abstraction blueprint.
- An ontology has descriptions, semantic relation titles, and typed properties; a schema is just field names.
- Integration = persistent, related, governed data. MCP connector = live, on-demand, unmodeled data. Use both, for different data.