All Features Compute Pools & Cost Reporting

Pick your clouds.
See what they cost.

Build a compute pool from real, provider-native machines across seven clouds. Track token and compute usage per user, with budgets and quotas that keep spend predictable.

Pools are built from each connected credential’s live catalog — concrete SKUs like cx33 or PRO2-S with real vCPU, RAM, disk, and price
Project, user, and installation scopes, resolved in that order
Strategy (Balanced, Pack, Spread, Smallest fit) and exhaustion policy (Queue, Fail, Fallback chain) per pool
Nodes are reused when they fit and stay warm for 30 minutes
Settings → Usage tracks token spend by model and day against a monthly budget; Admin → Costs and Admin → Usage cover the whole installation

Read the guide

Project Infrastructure compute pool editor showing a pool with strategy Pack, exhaustion policy Fallback Chain, three sources and seven allowed instances, with allowed offerings from Hetzner, Scaleway, and Vultr listing vCPU, RAM, disk, location, and monthly price, and a sold-out Hetzner offering marked unavailable

Real machines, not size labels

A compute pool lists the actual provider-native offerings it is allowed to rent — cx33, PRO2-S, vc2-4c-8gb, and more — across Hetzner, Scaleway, Vultr, Infomaniak, DigitalOcean, UpCloud, and GCP, each with its real vCPU, RAM, disk, and price. Browse each provider’s full catalog, filter by provider, region, vCPU, RAM, or price, and add a machine type without leaving the page. You choose the strategy (Balanced, Pack, Spread, Smallest fit) and what happens when a provider runs out (Queue, Fail, Fallback chain).

Catalog filters for provider, region, minimum vCPU, minimum RAM, maximum price, and availability above the Add instances from catalog list, where Hetzner offerings are marked Already allowed, Region sold out, Catalog only with an Add button, or Not selected

Filter a provider’s live catalog and add the exact instance types you want in the pool

See why a node was chosen

Resource requirements resolve from task, trigger, skill, and agent profile down to project and platform defaults. The chat’s infrastructure panel shows the resolved requirement alongside the placement decision — which node won, and why.

Chat session details showing the requested vCPU, memory, and disk, the observed and configured hardware, the current compute pool with its strategy, and a saved placement decision explaining why a warm Hetzner node was reused

Every placement decision is explained, not just made

One fleet, every provider

The Nodes page lists every machine across every connected provider side by side, with health, CPU/memory/disk usage, and price. Follow-up work reuses a warm node instead of waiting on a fresh boot.

Nodes page listing Hetzner, Scaleway, and Vultr machines side by side with status, health, observed and configured hardware, offering price, CPU, memory, and disk usage, and the workspaces on each

One fleet view across every connected cloud provider

Token and compute usage, per user

Settings → Usage shows your AI usage by model and by day, against a monthly budget, for traffic routed through the SAM AI proxy. It also tracks your compute usage in node-hours against your quota. Admins get the installation-wide view — LLM spend by model, day, and user with a projection, plus per-user node usage — in Admin → Costs and Admin → Usage.

Settings → Usage page showing total cost, requests, and input and output tokens, a per-model breakdown, a daily cost trend, and budget controls with daily token limits and a monthly cost cap

Token usage by model and by day, against a monthly budget

Previous Comments as Context Next Event Streams & Triggers

Ready to try SAM?

Deploy on your own cloud in minutes. Open source, free to self-host.