merovingian

Merovingian OS — the Source. A declarative engine for human–AI partnerships: model an organization as a graph of purposes, and Merovingian projects each person a scoped Claude Code and Codex workspace and reconciles the graph like infrastructure.

Two operations, one graph:

• **build** — projection. Given a person, project the workspace they're entitled to: the

purposes they can see, the knowledge buckets they can read/own, the skills/tools/agent that load (their slice of the tenant library, materialized into each harness's native layout), and the row-level scope on sensitive data. build = projection of the global definition onto a scoped target (Purpose × Scope × Human).

• **deploy** — reconciliation. The desired state is a graph.yaml **plus a library/** of

first-party skill/agent prompts and tenant-wide workspace.md instructions, all in the tenant repo (ADRs 0012/0018); deploy plan diffs them against the live SurrealDB (content as short hashes), deploy apply converges — structure-only, idempotent, referrer-safe on delete.

Enforcement is the backend's job, not the build's: sensitive data carries record-level SurrealDB PERMISSIONS keyed on a scoped JWT identity. Generation ≠ enforcement — the build generates the scope; the database honors it.