Argonix

🏗️ Architecture · Isolation · Secrets · Operator K8s

Security & Architecture
The technical blueprint behind Argonix

How Argonix is built, deployed, isolated and operated — for platform engineers, security architects and CISOs who need real answers, not slides.

Deployment models

Choose the topology that matches your control plane.

☁️ Argonix Cloud (EU)

Fully managed in Frankfurt. Logical tenancy on shared infra with strict per-organisation row-level scoping. Backups, upgrades, monitoring handled by us.

🏢 Dedicated tenant

Single-tenant database + storage in your chosen EU region (Frankfurt, France, Nordics). VPC peering optional. Available on Enterprise.

🏠 Self-hosted

Helm chart + Kubernetes Operator (22 CRDs). Air-gapped supported. BYO Postgres, Redis, S3-compatible storage, LLM endpoint.

High-level architecture

┌──────────────────────────────────────────────────────────────────┐
│                    Browser · CLI · Terraform                     │
└──────────────────────────────┬───────────────────────────────────┘
                               │ TLS 1.3 · JWT · API tokens
┌──────────────────────────────▼───────────────────────────────────┐
│  Frontend (Vue 3 SPA)        │       API (Django 5 + DRF)        │
│  Static, CDN-cached          │  RBAC · audit log · webhooks      │
└──────────────────────────────┬───────────────────────────────────┘
                               │
        ┌──────────────────────┼──────────────────────┐
        ▼                      ▼                      ▼
  ┌──────────┐         ┌──────────────┐       ┌──────────────┐
  │ Postgres │         │ Celery queue │       │ Object store │
  │  + pgvec │         │  (Redis)     │       │  (S3-compat) │
  └──────────┘         └──────┬───────┘       └──────────────┘
                              │
              ┌───────────────┴────────────────┐
              ▼                                ▼
       ┌─────────────┐                ┌──────────────────┐
       │  Workers    │                │  Argos LLM       │
       │ (monitors,  │                │  (vLLM/Mistral/  │
       │  CSPM, AI)  │                │   OpenAI EU)     │
       └──────┬──────┘                └──────────────────┘
              │
              ▼
     ┌──────────────────┐
     │  50 connectors   │  GCP · AWS · Azure · K8s · Slack ·
     │  (outbound only) │  GitLab · Jira · Prometheus · ...
     └──────────────────┘

Multi-tenancy & isolation

🧱 Row-level scoping

Every model has an organization_id. Permissions are enforced server-side via dry-rest-permissions + a custom has_organization_role check. No cross-tenant reads are possible — verified by automated test suite.

🔒 Worker isolation

Celery tasks receive only the org context they need. AI agent tools cannot escape the organisation scope of the calling channel. Connectors are stored per-org, never global.

📨 Outbound-only connectors

Argonix never opens inbound ports into your infra. All 50 connector types initiate outbound HTTPS calls from our workers (or your workers if self-hosted) to your APIs.

🛂 RBAC

4 organisation roles (owner, admin, editor, viewer). Resource-level permissions on connectors, monitors, alert channels and Argos channels.

Secrets handling

🔐 Envelope encryption

Connector credentials encrypted at rest with AES-256-GCM. Data keys wrapped by a KMS key (Hetzner KMS / AWS KMS / GCP KMS / Vault depending on deployment).

📤 Never logged

Secret fields are masked in the API, never written to logs, never echoed back. UI shows •••• with rotate-only mutation.

🔄 Rotation-friendly

Update credentials without recreating the connector. CRD-driven workflow on Kubernetes: change the Secret, Operator rolls the cached client.

Kubernetes Operator & GitOps

Manage the whole platform as code — 22 custom resources, all reconciled.

apiVersion: argonix.io/v1alpha1
kind: Monitor
metadata:
  name: api-prod-uptime
  namespace: ops
spec:
  organizationRef: { name: acme }
  type: http
  url: https://api.example.com/healthz
  interval: 30s
  alertChannels: [pagerduty-on-call]
  remediation:
    workflowRef: { name: restart-api }
📦 Connector, Monitor, AlertChannel, ChatChannel
📦 StatusPage, MaintenanceWindow, Workflow, DetectionRule
📦 Budget, TestSuite, SecurityPolicy, KnowledgeBase… (22 total)

AI agent safety

🧰 Scoped tools

Argos tools are explicit Python functions registered per connector. The LLM cannot execute arbitrary code — only call registered tools with typed arguments validated server-side.

✋ Human-in-the-loop

Mutating actions (restart, scale, kill, push) require explicit confirmation by default. Auto-execution is opt-in per action and gated by RBAC.

📝 Prompt & tool log

Every prompt, tool call and response is logged in the audit trail. Replayable for compliance, debugging and red-teaming.

🚫 No training on your data

Managed LLM providers configured with training opt-out. Self-hosted lets you wire any local model (vLLM, Ollama) — your data never leaves your perimeter.

Container & supply-chain hardening

🐧 Minimal images

Distroless / Alpine base. Non-root user. Read-only root filesystem. No shells in production images.

📜 Signed SBOM

SBOM generated at build (Syft), images signed (Cosign), provenance attached (SLSA Level 3 target).

🔍 Daily CVE scanning

Trivy + Grype on every production image, every day. Critical CVEs trigger an internal incident with a 7-day patching SLA.

🔐 NetworkPolicies

Helm chart ships default-deny NetworkPolicies. Only the explicit east-west traffic between Argonix components is allowed.

Explore More

Architecture review with our team

30 minutes with an engineer who built Argonix — bring your hardest questions.