Every connector self-describes via a ConnectorManifest.
Multi-account and per-source rules are first-class. Each row below is a
node on your mesh; tier choice is how thick the strand.
The connectors most knowledge workers touch every day. Carabase treats them as first-class; multi-account is a primitive, not a retrofit.
Edge harvesters read directly from local apps on your Mac. They share retention + tier infrastructure with the cloud OAuth connectors; they just skip the network hop.
Health, listening, transactions. Time-series-shaped data that lives in substrate tables rather than the artifact graph — the mesh resolver hydrates referenced points at query time.
The iOS app exposes HealthKit samples, Photos asset metadata, Significant Locations, Screen Time, and a Shortcuts inbox. Every push is gated by the same retention + tier system as the desktop side.
Browser extensions (Safari / Chrome / Arc) push per-navigation events into a streaming connector. Default tier is metadata-only; promotion to Tier 0 fires when the destination matches your entity graph (repo allowlist, watched orgs, etc.).
* Tier 2 default; Tier 0 promotion via materialization policy
(matches user's entity graph, repo allowlist, etc.).
** Beeper defaults to metadata-only; promotion to full body
fires when sender resolves to a known concept_role.
The mesh is the data plane. The engine runs over it.
Read the architecture →