Skip to main content

Architecture

Canonical public overview of how Nora’s control plane is wired together — frontends, API, queues, workers, backend adapters, and the runtime contract package.
Nora is the self-hosted AI agent ops platform — an operator-facing control plane for OpenClaw, Hermes, and supported sandboxed runtimes. Three browser surfaces sit behind one nginx ingress, platform state lives in PostgreSQL, background work is coordinated through Redis and BullMQ, and runtimes are provisioned (or proxied) through backend adapters that share the runtime contract package in agent-runtime/. The public repo centers on a single control-plane host. Agent workloads can stay on the local Docker host, be placed onto GA Kubernetes targets, use an experimental operator-registered Remote Docker host over SSH, or opt into experimental Proxmox LXC placement without changing the core operator workflow. NemoClaw is an experimental sandbox profile layered onto supported targets; NemoClaw on Proxmox remains blocked. Operator dashboard — the surface this architecture renders to the user

System map

Major components

Control plane

Request routing

WebSocket upgrades are handled at the nginx layer; the Express app attaches WS handlers via attachGatewayWS. SSE chat endpoints (/api/agents/*/gateway/chat) have chunked transfer encoding disabled for real-time streaming. OAuth callbacks land at the marketing app before redirecting into backend-issued HttpOnly session cookies.

API responsibilities

backend-api/server.ts is the control-plane integration hub. It wires together:
  • Security middleware: helmet, CORS, rate limiting, correlation IDs, prototype-pollution-safe input validation
  • Auth middleware (session JWT + workspace-scoped API key bearer auth)
  • Workspace-scoped route families for agents, alert rules, API keys, channels, integrations, LLM providers, monitoring, Agent Hub, and workspace members
  • Skill catalogs for both runtime families: ClawHub browse and install for OpenClaw agents, Hermes skills browse/search/install served from a disk-cached copy of the aggregated Hermes skills index, and the instance-curated Hermes Skills Library of pinned registry refs surfaced first in skill pickers
  • Session-only Remote Docker registration for personal hosts plus workspace-based host grants
  • Admin-only routes for platform Remote Docker hosts, fleet operations, audit, moderation, release upgrades, and platform settings
  • Gateway proxy with SSRF guards (gatewayProxy.ts — exposes createGatewayRouter + attachGatewayWS)
  • Background telemetry (backgroundTasks.ts, scheduler.ts) and release-availability checks (releaseInfo.ts)

State and queue boundaries

The API persists desired state first, then hands long-running work to a queue-backed worker. That keeps provisioning failures, retries, and delayed readiness out of the synchronous browser request path.

Workspace RBAC

Shared agents and other workspace resources use workspace RBAC. Personal Remote Docker hosts remain directly user-owned and may be granted into a workspace without transferring credentials or ownership. Platform hosts are owned by the control plane, managed by any platform admin, and may also grant access through workspaces, direct users, user groups, or all accounts. A workspace has a creator (workspaces.user_id) plus a row per member in workspace_members with one of four roles: Permission checks read from workspace_members.role, never from workspaces.user_id directly. Invitations land in workspace_invitations with a hashed token signed by NORA_WORKSPACE_INVITE_SECRET (falls back to JWT_SECRET). For either host type, a workspace viewer can see the host and an editor or higher can deploy. Workspace access never permits connection or credential changes: the personal owner manages a personal host, while any platform admin manages a platform host.

API keys and bearer auth

Each workspace can mint scoped API keys (/app/workspaces/:id/api-keys). Tokens are bearer-only, prefixed nora_, hashed at rest with HMAC-SHA256, and carry a fixed scope set: agents:read, agents:write, workspaces:read, monitoring:read, integrations:read, integrations:write. Workspace mutation, member management, key issuance, and every /remote-hosts management operation stay on session auth. Migration upload, draft, and live-inspection routes are also session-only. An API key cannot mint another key, manage a Remote Docker host, or ask the control plane to inspect a container or network source.

Migration and filesystem contract

Migration flow

Nora ships a control-plane-managed migration path for both openclaw and hermes:
  1. The operator prepares a migration draft from one of:
    • An uploaded Nora migration bundle or legacy OpenClaw template JSON
    • A local live Docker source, available only to a platform admin on a self-hosted control plane
  2. backend-api/agentMigrations.ts normalizes the imported data and stores an encrypted manifest in PostgreSQL.
  3. The operator deploys a new Nora-managed agent using that draft.
  4. The provisioning worker recreates the runtime under Nora control instead of adopting the original runtime in place.
Nora does not bind to the legacy runtime as-is. It uses the draft as desired state for a fresh Nora-managed deployment. Live SSH pull is disabled until Nora can require independently verified host-key trust. Export a bundle on the source machine and upload it instead. Live Docker pull is privileged because it reads through the control plane’s Docker socket; hosted deployments and non-admin users cannot invoke it.

Import surface

The current public import contract is intentionally scoped:
  • openclaw: agent files, workspace content, session memory, and provider material Nora can extract from supported source files
  • hermes: workspace content, model config, supported Hermes channel config, and provider environment material
  • Both families: supported Nora-managed state such as imported provider records, channel/integration wiring where available, and per-agent secret overrides
Unsupported runtime-specific state is surfaced as draft warnings instead of being silently invented or applied.

Export and live files

  • Nora-managed agents can be exported as nora-migration-bundle/v1 bundles for recreation on another Nora control plane.
  • The agent detail Files tab reads the actual runtime filesystem through backend-api/agentFiles.ts.
  • Filesystem access is root-allowlisted: a writable workspace root and curated read-only system roots for inspection and download.
  • The browser never receives arbitrary host filesystem access; reads and writes are mediated through runtime-aware backend commands.

Runtime provisioning

Selection model

Nora chooses a concrete backend through four layers of intent: The worker resolves the final backend through shared metadata in agent-runtime/lib/backendCatalog.ts. Placement is explicit: selecting remote:build-host does not let Nora substitute another host. The port allocator prevents Nora-managed published-port collisions, but Remote Docker has no aggregate host-capacity admission, automatic placement, failover, drain, or rebalance. See Provisioner backends for the supported combinations and their maturity.

Provisioning lifecycle

Bootstrap contract

The worker and backend adapters share one runtime bootstrap package from agent-runtime/:
  • Runtime library files are injected into the launched environment.
  • Template payload files are copied into the runtime workspace when present.
  • Runtime environment variables are assembled centrally.
  • Endpoint conventions stay shared across the worker and API.
Shared ports:

Backend adapter code sharing

Compose mounts ./workers/provisioner/backends into backend-api at /app/backends. Adapter code is physically shared between the API and the provisioner worker. When editing adapters, verify both consumers still work.

Background work

Failed jobs are retained on each queue for inspection. The DLQ surface for deployments is exposed via getDLQJobs / retryDLQJob in backend-api/redisQueue.ts.

Trust boundaries

  • Browsers never talk directly to PostgreSQL or Redis. All browser traffic enters through nginx and reaches stateful services through the frontends or backend-api/.
  • backend-api/ owns auth, persistence, queue orchestration, release metadata, and runtime-facing proxy routes. Frontends do not provision runtimes directly.
  • backend-api/ also owns migration draft inspection/storage and all runtime file access mediation. Browser users do not receive direct host or container filesystem access.
  • workers/provisioner/ and workers/backup/ handle long-running infrastructure work outside the request path. They consume queued jobs and write the result back into control-plane state.
  • agent-runtime/ defines the runtime-side contract used after launch. Control-plane code depends on that contract rather than embedding backend-specific assumptions everywhere.
  • External execution systems such as Docker, experimental Remote Docker over SSH, Kubernetes, experimental Proxmox LXC placement, and NVIDIA secure sandboxes are reached through backend adapters instead of directly from browser surfaces.
  • Remote-host SSH credentials are encrypted at rest and masked in API responses. Only the personal owner or a platform admin for the corresponding host type can replace them; all-account, direct-user, user-group, and workspace grants permit placement only and never reveal credentials.
  • Stored secrets — provider keys, integration credentials, OAuth tokens, backup encryption material, SMTP password — are AES-256-GCM encrypted with ENCRYPTION_KEY. API keys and Agent Hub installation keys are HMAC-hashed at rest, not encrypted.

Deployment topologies

Fleet view — control plane managing multiple agent placements The clearest public path today is one host running the control plane, with agent runtimes launched locally through Docker by default. Public-domain setups can either let Nora own public ingress directly or put an external reverse proxy in front of Nora’s internal nginx.

Current constraints

  • The public OSS path is primarily a single-host control plane. The repo does not currently claim a first-class HA or distributed control-plane deployment story.
  • OpenClaw is the default runtime family. Hermes is a narrower, deployment-first runtime path with a different operator contract.
  • Migration recreates runtimes under Nora control; it does not adopt a legacy runtime in place.
  • Hermes is a runtime family, not a backend id. Docker and Kubernetes are the current GA Hermes execution targets. Remote Docker Hermes support is experimental and currently requires the agent deployment API because the Hermes picker does not merge registered hosts.
  • Kubernetes (k3s/k8s) is a GA execution-target option for agents. Remote Docker, NemoClaw, and Proxmox LXC placement are experimental.
  • Remote Docker’s manual connection result has no expiry or continuous health loop. Nora does not automatically move workloads when a remote host becomes stale, unreachable, or overloaded.

Provisioner backends

Supported deploy targets, their maturity, and required configuration.

Environment variables

Every variable Nora reads, grouped by subsystem.

Workspaces

Workspace ownership, RBAC roles, member invitations, and Remote Docker grants.

Remote Docker

SSH, test scope, explicit placement, sharing, capacity limits, and safe retirement.