Monarch
Agents propose. They never dispose.
Monarch's security model starts from a structural premise, not a policy document: no agent in the system holds the authority to take a consequential action on its own. Authority lives in one governed boundary — and everything that crosses it is recorded.
One gate for every consequential action
When an agent wants to send, write, purchase, or change anything outside the platform, it files an execution request. Gatekeeper — a dedicated service, not a code path inside the agent — decides what happens next.
Each request carries the acting agent, initiating user, organization, provider, action type, full payload, business context, and an assessed risk level — enough for a human or a policy to judge it on its face.
Per-action approval settings decide what may run unattended. Routine, low-risk actions on trusted connections clear automatically — with the policy result recorded alongside the action.
Consequential actions wait for a person. Sign-off cards in the workspace show exactly what will happen before it happens — approve, deny, or unblock, with reviewer identity and timestamp recorded.
Action settings can restrict approval and use to specific organization roles — governance follows your org chart, not just your credentials.
Approved actions execute through managed connectors or vetted direct integrations using approved secrets — never from inside an agent's reasoning loop. Idempotency keys prevent double execution.
Results, errors, reviewer, and timing are written back to the request. The question "what has this system ever done on our behalf?" has a complete, queryable answer.
Isolation by construction, not by convention
Organization-scoped everything
Every tenant record — jobs, documents, connections, actions, chats — is keyed to an organization and protected by row-level security policies that preserve isolation at the database layer.
Multi-organization users
People belong to multiple organizations with distinct roles in each. Every flow resolves the active organization explicitly before creating or reading anything — no accidental cross-tenant work.
Server-side sessions only
Enterprise-grade identity handles sign-in; sessions live server-side. The browser talks to same-origin APIs and nothing else — no tokens, no internal service URLs, no authorization plumbing in frontend code. That rule is architectural, and it's enforced.
An append-only memory of everything that matters
Monarch treats audit as a first-class data product, not a log file. Multiple dedicated streams capture activity at the fidelity compliance work actually needs.
- User action history with integrity hashing. Administrative, data, connection, and agent activity — actor, target, category, severity, outcome, source IP, route, status, and duration — in an append-only record designed to resist tampering.
- The job event stream. Every plan, task, run, and execution request emits events as it moves — the same stream that powers live progress is your forensic timeline.
- Connection & admin events. Configuration changes and outbound sends on administrative connections are logged with actor and result.
- Acknowledged notifications. Required notices track per-user read and acknowledgement state — when you need proof that everyone saw the policy change, you have it.
Scanned as part of shipping
Deployed services are exercised by automated dynamic security scans on a standards-aligned assessment profile, wired into the delivery pipeline as a gate. Every run — scanner version, target, findings, exit status — is recorded in the platform's own history.
No secrets in code, ever
Credentials load through managed secret references at deploy time — never committed, never passed as command arguments, never shipped to the browser. Webhook signing keys are stored encrypted, and encrypted storage is validated with round-trip checks in the platform's own test suite.
Defense in depth, summarized
Agents can't act directly — authority is architecturally concentrated in one reviewable boundary.
Per-action approval modes, risk levels, and role restrictions govern what runs and who decides.
Tenant-keyed records, row-level security, and server-side identity keep organizations apart.
Append-only audit, integrity hashing, recorded scans, and a complete action ledger.
Governance that doesn't slow you down.
Approval where it matters, autonomy where it's safe — see what that feels like in practice.