elwin@sytadel:~$ ▋

sytadel-suite

Security and identity control plane for teams building with AI agents.

# nota: Actively developed product: the repo holds working services and prototypes, not a production-ready hosted runtime. Scale metrics don't apply yet — the honest story is the security model and the architecture.

## problem

Teams building software and automations with AI agents have to rebuild trust infrastructure from scratch: identity and multi-tenancy, policy enforcement, secrets and vault, tamper-evident audit, agent identity, MCP and human approval of sensitive actions.

  • It's expensive security work, easy to get wrong, and a distraction from the real business domain.
  • Getting it wrong stays invisible until it's too late: a cross-tenant leak, an altered log, an exposed secret.

## architecture

A monorepo of independently deployable services orchestrated with docker-compose. Every request is authenticated at the edge, authorized against per-tenant policy, and only reaches the vault as a signed service-to-service call.

  • auth-api — identity, tenants and memberships; the single source of truth of the identity plane.
  • zerotrust-api — enforcement gateway: validates the JWT, resolves per-tenant policy and signs the downstream call.
  • securechain-vault — documents, encryption, tamper-evident audit and notary.
  • billing-api — plans, metering and entitlement sync.
  • sytadel-mcp-server — exposes typed tools over the identity plane for MCP clients.
  • No cross-service foreign keys: services resolve tenants/memberships against auth-api's internal directory.

## technical decisions

  • Tamper-evident audit: SHA-256 hash chain, append-only enforced at the DB (trigger + REVOKE), concurrency-safe (pg_advisory_xact_lock), and anti-truncation checkpoints anchored with RFC 3161 timestamps.
  • Document notarization: a real Merkle tree (RFC 6962) + RFC 3161 timestamps, with an honest SIMULATED fallback when no TSA is configured — never a fake proof.
  • Encryption at rest: AES-256-GCM envelope encryption, per-tenant data keys (DEK) and a random IV per operation.
  • Zero-trust service-to-service: no service trusts an unauthenticated or unsigned request; identity is re-derived from a verified JWT and carried in a signed canonical request.
  • AI as a policy compiler: an LLM translates natural language → RBAC, reviews access and classifies session anomalies — all with evals (precision/recall, cost, latency).
  • MCP + human-in-the-loop: a LangGraph agent proposes access decisions that a person applies.

## results

The project's guiding rule: the MVP may be incomplete, but it must not violate tenant isolation, access control, document integrity or auditability.

  • Tenant isolation enforced server-side from the verified token, never from client-supplied input.
  • Smoke tests that walk the auth → zerotrust → vault flow, including a negative case (creating a tenant in the vault returns 409 because that's owned by auth-api).
  • The shared base that made it possible to extract @elwinernst/nest-auth, the npm package for JWT verification shared across services.