Which VPS Can Handle Your Agent’s Busiest Job?
Measure a cold-start burst on each orderable VPS, then compare monthly infrastructure cost among plans that meet your memory and completion-time targets. If you run Claude Code, apply its documented memory floor as well.
Size from whole-host measurements
Start with the memory requirement for the software you actually install. Claude Code lists 4 GB or more of RAM and supports x64 and ARM64. On each candidate VPS, use the same agent version, repository snapshot, and job fixture. Reboot the pilot, let the gateway settle, then start your expected peak number of simultaneous jobs. Record whole-host peak used memory from a consistent measurement, plus the time until the last job finishes. Include the browser, containers, and any indexing processes in the run. Repeat the burst three times and retain the highest memory and completion-time readings for each plan. An illustrative host with 8 GiB usable and a 6.9 GiB peak leaves 1.1 GiB free, so it misses a 1.5 GiB reserve even if the jobs complete. Check for swap growth or killed processes before accepting a fast result.
Advanced setup - Claude Code DocsPut the replacement host in the pilot
OpenClaw’s VPS guide places gateway state and the workspace on the host. That makes replacement a separate test from throughput. Seed the pilot with a synthetic workspace file and one harmless scheduled job. Restore an approved backup to a fresh instance in the same region, using a test channel or an isolated route. Keep the original schedule fenced from the restored copy so a single scheduled event cannot fire on both. Record minutes until the replacement can read the file and produce one expected job result, then check the workspace against its saved inventory. When comparing quotes, use the same currency and billing basis, and itemize compute, charged backups or volumes, and expected egress. Record the quote date and region yourself. A cheap compute line can lose once the backup needed for your measured recovery target is included.
Linux server - OpenClawSet the floor from the installed tools
List everything that will run on the VPS at the same time, including the agent, its tool processes, and any browser or build jobs. Claude Code documents a minimum of 4 GB of RAM on x64 or ARM64. If you plan to install it on the server, use that as the starting point, then verify that each dependency works on the chosen architecture. In particular, test the actual browser, container images, or native packages you expect to use.
For an illustrative sizing trial, assume one Claude Code session, a repository checkout, and one test runner. Start with a 2 vCPU, 4 GB candidate and run a complete job, including the busiest expected test phase. Watch memory, swap, disk use, and process restarts. A job that finishes after an out-of-memory restart still fails the sizing trial if it misses your recovery or response target.
Advanced setup - Claude Code DocsCompare plans you can order in your region
OpenClaw's Linux server guide lists DigitalOcean as a paid VPS option and Oracle Cloud as an Always Free ARM option. For DigitalOcean and Vultr, check each provider's order screen for the same region or latency target, CPU architecture, RAM, disk, and billing term. For Hetzner, check availability for the exact SKU and location. For Oracle, confirm you can actually provision the ARM shape and review its current eligibility, capacity, and reclamation terms before depending on it.
For an illustrative budget, assume a quote of $18 per month for compute, $2 for an address, $4 for backups, and $11 of separately metered model use. The total would be $35 per month before tax or extra traffic. Apply the same calculation to Linode, Lightsail, Contabo, or a managed quote if they are on your shortlist. Verify each amount at checkout and check how partial months are billed.
Linux server - OpenClawMake replacement part of the pilot
OpenClaw's VPS guide places the Gateway, state, and workspace on the server. Deploy a test Gateway using the project's Linux server guide, connect one test channel, and send a message that requires reading a known workspace file. Save the config and workspace backup somewhere independent of the server. Keep credentials in your approved secret store and record how a replacement machine receives them.
Reboot the test VPS and send the same request again. Then use an isolated replacement to restore the backup and verify that the workspace file and intended channel access both return. Record the elapsed recovery time and the age of the restored data. A working Gateway with an empty workspace indicates a backup or mount problem. A restored file with a silent channel calls for a separate check of credentials, startup logs, and message delivery.
Linux server - OpenClawFind the limit before buying a larger VPS
OpenClaw's hosting FAQ describes the tradeoff between a headless VPS and a local machine with a visible browser. If an agent fails only during an interactive browser step, reproduce that step in the deployment environment before upgrading the server. Check whether the browser can launch and complete the interaction without a visible desktop. If the task requires a visible window, compare local execution on a machine that can stay awake with a hosted option that explicitly supports the interaction.
If simple text tasks fail too, line up the job log with server metrics and provider responses. An out-of-memory kill during a test run supports reducing concurrency or adding RAM. An Anthropic HTTP 429 rate-limit error calls for checking the relevant subscription window or API quota. Repeat the same failed job after changing one variable and confirm that the original symptom is gone.
FAQ: providers, hardware, and hosting - OpenClaw