Resource requirements
CPU, memory, and storage for Exo workloads.
We run the Exo control plane. Your cluster runs the Operator, Agents, and any gateways or tools you enable.
Resources
These are defaults for the Operator and Agent runtime. You can override them in Operator Helm values or Agent resource settings. Requests reserve capacity, while limits cap usage. 1000m equals 1 vCPU.
| Component | CPU request / limit | Memory request / limit |
|---|---|---|
| Operator | 100m / 500m | 128Mi / 512Mi |
| Agents | 500m / 1 vCPU | 512Mi / 1Gi |
| Lightweight Agents | 100m / 250m | 50Mi / 100Mi |
Plan about 5 vCPU and 10 GiB for ten concurrent Agents using the standard runtime. Add room for the Operator, Kubernetes, tools, upgrades, and node failures. Check actual usage before scaling up.
Measured usage
We measured about 2.8 vCPU and 3 GiB across five OpenCode Agents in short lab tasks. The idle Operator used only a few millicores. Larger repositories, builds, and more sessions can need more resources.
Storage
| Storage | Starting amount | How to count it |
|---|---|---|
| Image cache | About 1 GiB per worker node | Runtime/addon images shared on that node. More images need more space. |
| Temporary Agent files | About 1 GiB per running Agent | Writable files and workspace. Builds may need more. No PVC. |
| AgentFS PVC | 5 GiB per enabled volume | Tested default, configurable. |
| Persistent workspace PVC | 1 GiB per enabled workspace | Tested default, configurable. |
For five running Agents on one node, start at about 6 GiB before Kubernetes, logs, upgrades, and PVCs. The 1 GiB figures are planning assumptions, not measured limits. PVC requests are capacity, not files already saved.