Where should each OpenClaw Gateway run?

Start with a Linux VPS and private admin access. Use Docker Compose when you want an image-based rollout, and check your Gateway inventory for shared trust boundaries before onboarding users.

Assign state ownership before choosing a host

Treat the Gateway's state and workspace as the assets you have to recover. For each deployment, record the owner, the backup location, and the person who can reach the Control UI through Tailscale or SSH if the primary operator is unavailable. A Linux VPS gives your team direct responsibility for those pieces. For a managed option, ask the operator to demonstrate an export and a restore into a fresh instance. Run the same drill on both candidates with a disposable workspace containing a known text file. From a second operator's device, locate that file after restore and time the first successful Control UI reply. A healthy dashboard with an empty workspace is a failed restore. Compare elapsed minutes and list every step that required the original operator or hosting support.

Linux server - OpenClaw

Check rollout identity and isolation independently

Draw one row per trust boundary, with a Gateway ID, host, and OS account. Separate Gateway IDs let you distinguish deployments, while distinct OS accounts or hosts provide a stronger separation between mutually untrusted groups. Docker packaging and agent sandboxing are separate choices. The Docker guide says sandboxing starts disabled, so record its actual configuration if your acceptance criteria depend on it. For an OCI rollout, record the deployed manifest digest and the previous known-good digest in your change record. A pilot that builds from source needs at least 6 GB RAM for that build according to the guide. A prebuilt image avoids that build requirement. Before expanding the pilot, compare the digest recorded at approval with the one actually deployed, and verify the intended test workspace from each Gateway.

Docker - OpenClaw

Decide who can share a Gateway

OpenClaw's Docker guide recommends a separate Gateway for each trust boundary, ideally under separate operating system users or hosts. Start by listing who may message the bot, which credentials it needs, and whose files it can read. Use those answers to decide where to provision a separate Gateway. OpenClaw also documents that sandboxing is off by default, so verify the settings your deployment actually uses.

For an illustrative rollout, assume two unrelated clients, each with a private messaging channel and a different model credential. Give each its own Gateway, channel identity, and test workspace. Put a distinct harmless marker in each workspace. A pass means each approved sender can retrieve their own marker and the operator can confirm that only the matching workspace and credential are attached to that Gateway.

Docker - OpenClaw

Bring up one Linux Gateway

The OpenClaw Linux hosting guide describes the VPS as the home of both Gateway state and the workspace. It lists DigitalOcean as a VPS option and points to a Docker deployment on Hetzner. Choose one host and record where the persistent state and workspace will live. For the Compose route, prepare Docker Engine and Compose v2, get the official OpenClaw repository, and follow its Docker setup guide from the repository root.

Configure a model credential and one test messaging channel during onboarding or through the relevant configuration flow. Send a uniquely identifiable test message from an approved account and confirm a reply in the same channel. Restart the Gateway and repeat the test. If it responds only before the restart, confirm that the restarted service reads the intended configuration and workspace.

Linux server - OpenClaw

Separate channel failures from model failures

When the bot goes quiet, check the Gateway or provider logs for the test request before changing the deployment. If the request never reached the Gateway, verify the channel configuration, sender access, and bot identity. If the request reached the model and returned Anthropic's rate_limit_error with HTTP 429, OpenClaw's FAQ points to a provider quota or rate limit for the current window.

For an Anthropic API key, inspect usage and limits in Anthropic Console. Retry one small message after the limit resets or after correcting the quota. If you configure a fallback model, test it with harmless input and verify the provider that served the reply. If the error mentions extra usage for long context requests, check the credential's long-context billing eligibility before changing channel settings.

FAQ: providers, hardware, and hosting - OpenClaw

Sources