Documentation Index
Fetch the complete documentation index at: https://docs.nusomi.com/llms.txt
Use this file to discover all available pages before exploring further.
Self-hosted Nusomi runs the control plane, capture ingest, indexing, replay, and storage inside your network. Frames and events never reach Nusomi-managed infrastructure.
When to self-host
You probably want self-hosted if any of these are true:
- Your workflows touch regulated data (PHI, PCI, regulated financial data) and your security team has a hard “no third-party processors” rule.
- You’re in a sovereignty regime (data must remain in a specific country / jurisdiction).
- You already operate a private model gateway and want training data to flow into it without leaving your network.
If none of these are true, the SaaS deployment is simpler. The two share a wire protocol — code written for one runs on the other unchanged.
Architecture
┌───────────────────────────────────────┐
│ Your VPC │
│ │
│ ┌─────────────┐ ┌──────────────┐ │
│ │ Desktop │ │ Browser │ │
│ │ agents │ │ extension │ │
│ └──────┬──────┘ └──────┬───────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌─────────────────────────────────┐ │
│ │ Nusomi control plane │ │
│ │ (auth, scheduling, dashboards) │ │
│ └──────────────┬──────────────────┘ │
│ │ │
│ ┌──────────────┼─────────────────┐ │
│ ▼ ▼ ▼ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Ingest │ │ Indexer │ │ Replay │ │
│ │ workers │ │ workers │ │ workers │ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ ▼ ▼ ▼ │
│ ┌────────────────────────────────────┐ │
│ │ Postgres · Object store · Redis │ │
│ └────────────────────────────────────┘ │
└───────────────────────────────────────┘
Components
| Component | Notes |
|---|
| Control plane | Stateless API + dashboard. Horizontally scalable. |
| Ingest workers | Receive frames + events, persist to object store + Postgres. |
| Indexer workers | Build memory graph, derive sealed-session metadata. |
| Replay workers | Run exact / guided / partial replays. Talk to your model gateway. |
| Postgres | Sessions, events, memory graph (≥ 14). |
| Object store | Frames. S3, GCS, MinIO, or Ceph compatible. |
| Redis | Streaming, queues, rate limits. |
| Model gateway | Yours. We support any OpenAI-compatible / Anthropic-compatible endpoint. |
| Format | Notes |
|---|
| Helm chart | Default. Targets EKS, GKE, AKS, OpenShift, plain Kubernetes ≥ 1.27. |
| Docker Compose | Single-host development / pilot. |
| Air-gapped tarball | Includes mirrored container images, signed manifests, no internet access required. |
Updates
| Cadence | Notes |
|---|
| Patch (security + bugs) | Bi-weekly |
| Minor (features, additive APIs) | Monthly |
| Major (potentially breaking) | Quarterly, with 90-day deprecation window |
Each release ships:
- Container images (signed).
- A migration plan (ordering, downtime expectations, rollback notes).
- Release notes — see changelog.
Air-gapped customers receive a signed bundle on the same cadence.
Connectivity to nusomi.com
Self-hosted is self-contained. The only outbound calls are:
- License verification, once per 24 hours, to
license.nusomi.com. Air-gapped deployments use an offline license.
- Optional anonymous telemetry (off by default).
- Optional model gateway (yours).
Frames, events, and metadata never leave your network.
Sizing
A starting point for a 100-engineer workspace, ~50k captured minutes per month:
| Tier | Nodes |
|---|
| Control plane | 3 × c-medium |
| Ingest workers | 4 × c-large |
| Indexer workers | 2 × c-large |
| Replay workers | 2 × c-large (more if replays are continuous) |
| Postgres | 1 × db-large (HA replica) |
| Object store | 5 TB to start, growing ~1 TB / month / 10k captured minutes |
| Redis | 1 × r-medium |
The Helm chart’s values.yaml ships with this profile under profile: starter. Larger profiles (mid, large, xl) are documented inline.
Authentication
The control plane fronts:
- SAML / OIDC SSO for dashboard.
- OAuth 2.0 client credentials for service-to-service.
- API keys minted by your control plane (
nsk_local_...).
Keys never reach Nusomi infrastructure.
Backup & DR
- Postgres: point-in-time recovery via your provider’s standard tools.
- Object store: versioned, cross-region replication.
- Redis: ephemeral; lost on restart (queues will re-form).
A reference DR runbook ships with the chart.
Next