On-premise · reference
System requirements
What you need to run DashboardGPT on your own servers. Use these as starting points: we confirm final sizing with you on the scoping call, based on users, data volume and model choice.
Last updated:
Application servers
For the DashboardGPT app, workers, PostgreSQL and Redis. Excludes local LLM hardware, which is listed separately below.
| Footprint | Evaluation | Standard | Large / HA |
|---|---|---|---|
| Typical use | Pilot, up to 10 users | Up to 50 active users | 50+ users, high concurrency |
| Install method | Docker Compose | Docker Compose or Kubernetes | Kubernetes with Helm |
| CPU | 4 vCPU | 8 vCPU | 16+ vCPU across 3+ nodes |
| Memory | 16 GB | 32 GB | 64+ GB across nodes |
| Disk | 50 GB SSD | 100 GB SSD | 200+ GB SSD, plus managed DB storage |
| PostgreSQL | Bundled container | Bundled or external | External, with replication |
| Redis | Bundled container | Bundled or external | External or Sentinel / cluster |
Add storage for uploaded CSV and Excel files and for cached query results if your teams upload large files.
Software
| Component | Supported versions | Notes |
|---|---|---|
| Operating system | Ubuntu 22.04 / 24.04 LTS, RHEL 8 / 9, Rocky / Alma Linux 8 / 9 | Any 64-bit Linux that runs a current container runtime |
| CPU architecture | x86_64, arm64 | Multi-arch images |
| Docker | Docker Engine 24+ with Compose v2 | For single-host installs |
| Kubernetes | 1.27+ · Helm 3.12+ | EKS, AKS, GKE, OpenShift 4.14+, RKE2, k3s |
| Metadata database | PostgreSQL 14+ | Stores users, canvases, settings and the audit log |
| Cache and queue | Redis 7+ | Valkey 7.2+ also supported |
| TLS / ingress | NGINX, Traefik, HAProxy or any Kubernetes ingress controller | Bring your own certificate |
| Email (optional) | Any SMTP server | Invitations, alerts and scheduled reports |
| Identity (optional) | SAML 2.0, OpenID Connect, LDAP / Active Directory | Okta, Entra ID, Keycloak, Ping, Google Workspace |
Language model
Choose one. You can switch later, or use different models in different workspaces.
- Option A · no GPU needed
Bring your own key
The app calls your existing provider. You pay the provider directly at your negotiated rates.
- OpenAI, Anthropic, Azure OpenAI or Google Gemini
- AWS Bedrock and others through an OpenAI-compatible gateway such as LiteLLM
- Outbound HTTPS to the provider endpoint only
- Option B · fully offline
Local model
Run an open-weight model on your own GPUs and point DashboardGPT at it.
- Ollama, vLLM, TGI, LM Studio or any OpenAI-compatible API
- No outbound network access required
| Local model size | Example GPU setup | Good for |
|---|---|---|
| 7–14B parameters | 1 × GPU with 24 GB VRAM (L4, A10, RTX 4090) | Pilots, simple schemas, lower concurrency |
| 30–35B parameters | 1 × 48 GB (L40S, A6000) or 2 × 24 GB | Most teams. Good SQL accuracy on multi-table schemas |
| 70B+ parameters | 2 × 80 GB (A100, H100) or 4 × 48 GB | Complex schemas, many concurrent users |
GPU figures assume 4-bit or 8-bit quantised weights served by vLLM or Ollama. Answer quality depends on the model you choose. We share our current recommended models on the scoping call.
Network and ports
Inbound and internal traffic is all that an air-gapped install needs.
| Direction | Port | Purpose | Air-gapped |
|---|---|---|---|
| Inbound | 443 | Users reach the web app through your load balancer | Required |
| Internal | 5432 · 6379 | App to PostgreSQL and Redis | Required |
| To your data | Database port | App to your warehouses, e.g. 5432, 1433, 1521, 443 | Required |
| To local LLM | 8000 / 11434 | App to vLLM or Ollama endpoint | If using a local model |
| Outbound | 443 | Your cloud LLM provider (BYOK) | Not used |
| Outbound | 443 | License check and image registry | Not used, offline license |
Clients and operations
- Browsers
- Latest two versions of Chrome, Edge, Firefox and Safari. The infinite canvas works best on screens 1280 px and wider.
- Backups
- Back up the PostgreSQL metadata database and the uploads volume. Your source databases are only read, never written.
Get a sizing recommendation for your environment.
Share your user count, data sources and model preference. Our engineers will send back a sized architecture before you install anything.
- 30-minute call with a solutions engineer
- Architecture and sizing review
- Trial license for your own servers