PROJECT TYPES
A project in Talos has a type, and the type decides which areas it gets: an AWS project gets inventory and stacks, a Proxy project gets routes and pools. Everything underneath is shared: the same queue, the same permissions, the same audit ledger. This page maps each type to the pages that describe it, and says plainly how proven each one is.
Accounts, every region’s inventory, EC2, S3, RDS, Route 53, ACM, monitoring, cost, Lambda, ECR and EKS, plus resource graph, posture checks, fleet operations and repository-driven Terraform.
Graph, posture, fleet operations and IaC run against a real sandbox account.
Service-principal accounts, resource-group inventory, VMs, blob storage, stacks, Flexible Server, DNS, Monitor, cost, Functions and ACR.
Service-account access, inventory, Compute Engine, Cloud Storage, stacks, Cloud SQL, Cloud DNS, monitoring, billing-export cost, Cloud Functions and Artifact Registry.
OCI compute, object storage, stacks, databases, DNS, cost, Functions, OCIR and OKE, plus the Fusion Applications estate: environments, refreshes, ESS jobs, SCIM and BI Publisher.
Clusters, namespaces, workloads, Helm releases and per-job credentials, whether self-run or bridged in from a cloud.
Proxmox and libvirt nodes, VMs, consoles, snapshots, clones, storage, ISOs, networks, backups and migrations.
Not yet run against a live Proxmox or libvirt host.
nginx and HAProxy routes and pools, traffic splitting, rate limits, access control, WAF, fail2ban, Squid egress and keepalived pairs.
Not yet run against a live nginx or HAProxy host.
Zones and records across five clouds and self-hosted nameservers, domain expiry, ACME issuance and renewal, certificate deployment, DNSSEC, revocation and CT monitoring.
Not yet run against a real CA, registrar or nameserver.
Compose stacks, registries, containers and environment, deployed with health gates and planned before they are applied.
Application deployments from a repository to your servers, through reviewable pipelines.
GPU fleet, model registry, training runs, gated promotion, canary serving, drift and cost.
Cloud GPU provisioning has not been run live.
MCP servers and hosted vendor connections, tool approval by hash, agents with enforced budgets, evaluations, and Talos exposed as an MCP server.
Not yet run against a live model provider or vendor consent screen.
Database instances, query console, users, migrations, backups with point-in-time recovery, restore verification and replication.
Bench environments, Frappe and ERPNext sites, custom apps, and site-level backup and restore.
SHARED BY EVERY TYPE
These are not project types of their own. They sit underneath all 14, which is why managing an EKS cluster and a Frappe site feels like the same product.
Hosts added over SSH and discovered read-only, for any project that uses them.
Every action runs as a queued, retryable job with live logs and a record of who ran it.
Multi-stage pipelines and two-person approval before anything irreversible.
Health checks, alerts, live logs, and a dashboard built for each type.
Role-based access with permissions scoped to named resources.
An AI assistant that reads the fleet and proposes changes for your approval.
Install it on a host you already have. Every type connects read-only first.