Guide · 2026-11-09

OpenClaw Enterprise Setup Checklist (2026)

Most OpenClaw rollouts fail in the boring places. Someone installs it on a laptop, connects one chat channel, and it works. Three weeks later the same instance is answering messages from a whole team, nobody knows where the config lives, and the one person who set it up is on holiday.

This checklist is for the second stage, when OpenClaw stops being a personal tool and becomes something a company depends on. It follows the order we would do the work in. Every OpenClaw fact below links to the official docs, and anything we could not confirm there is left out.

If you would rather skip the setup and get a running instance fast, AyClaw Bots deploys a managed OpenClaw instance from a model pick and a Telegram connection. The rest of this guide covers what you decide when you run it yourself.

What OpenClaw is, in one paragraph

OpenClaw is an open-source, self-hosted AI assistant under the MIT license. You run one Gateway process on your own machine or server. The Gateway connects chat apps to AI agents and is the source of truth for sessions, routing and channel connections. The docs list Discord, Google Chat, iMessage, Matrix, Microsoft Teams, Signal, Slack, Telegram, WhatsApp and Zalo as supported channels, with more through plugins.

That one-process design shapes the whole checklist. You are not deploying a fleet of services. You are looking after one Gateway, its config file, the credentials it holds and the tools it is allowed to run.

Phase 1. Decide the shape before you install

  1. Name an owner. One person owns the Gateway, its config and its credentials. A second person can restore it.
  2. Pick the trust boundary. The security docs say OpenClaw has one trust boundary per Gateway and assumes a single operator or a team whose members trust each other. It is not a hostile multi-tenant boundary. If two groups should not see each other's data, plan two Gateways.
  3. Choose the host. The install docs list VPS providers, Render, Kubernetes, a macOS VM, and containers through Docker or Podman. Pick the one your ops team already patches.
  4. List the channels. Start with one. Every extra channel is another place where strangers can send the agent text.
  5. List the tools you want the agent to have. Email, calendar, shell access and file access are different risk levels. Write down which agent gets which.
  6. Pick a model provider and budget. You need an API key from your provider. Decide who holds it and who sees the bill.

Phase 2. Install and verify

  1. Check the runtime. The install docs require Node 24 or 26 and recommend Node 26. They also advise newer patch versions for security, so check the current numbers on the install page rather than copying them from a guide.
  2. Run the installer or a package manager. The docs recommend the installer script on macOS, Linux and WSL2, and a PowerShell script on Windows. npm, pnpm and bun installs are also documented. Read any script before you pipe it into a shell, and pin a version for production instead of installing the latest tag.
  3. Run onboarding. It offers Quick start or Custom setup. Choose Custom for anything company-owned.
  4. Install the Gateway as a service. The docs use openclaw gateway install and openclaw gateway start. On Linux the service is a systemd user service, on macOS a LaunchAgent, and on native Windows a Scheduled Task.
  5. Verify. Run openclaw --version, openclaw doctor and openclaw gateway status. Save the output in your runbook.

Phase 3. Lock down the config

OpenClaw reads settings from ~/.openclaw/openclaw.json, written in JSON5, per the configuration docs. A few behaviors matter for a team.

  1. Treat the file as code. Keep a copy in version control with secrets removed. You can set a custom path with OPENCLAW_CONFIG_PATH.
  2. Know that the schema is strict. Unknown keys, wrong types or invalid values stop the Gateway from starting. If validation fails at boot, only openclaw doctor, openclaw logs, openclaw health and openclaw status work until you fix the file.
  3. Know how hot reload behaves. The Gateway applies valid edits automatically. An invalid edit during hot reload is skipped and the runtime keeps the last accepted config.
  4. Use doctor for repairs. openclaw doctor shows the exact issues and openclaw doctor --fix applies repairs. The docs say doctor is the only thing that restores the last-known-good config.
  5. Back up before changes. The docs describe a .bak ring for migration writes and a backup create command. Take a backup before every upgrade.

Phase 4. Connect channels and pair users

  1. Keep the defaults on at first. According to the security docs, a regular host install binds the Gateway to loopback, most chat channels answer unknown direct-message senders with a pairing code instead of processing the message, and group access is allowlisted behind a mention gate.
  2. Restrict who can talk to the agent. The docs recommend channels.whatsapp.allowFrom and group mention rules as a starting point. The same idea applies to every channel. Use an explicit sender list.
  3. Choose a direct-message access mode on purpose. The documented modes are pairing, allowlist, open and disabled. For a company instance, open is rarely right.
  4. Check how each channel authorizes people. The docs warn that some workspace channels trust workspace membership by default. Read each channel's page before you assume it is stricter.
  5. Test with a second account. Send a message from a person who is not on the allowlist and confirm the agent does not act on it.

Phase 5. Set tool permissions and sandboxing

This is where an enterprise setup differs most from a personal one.

  1. Turn on the sandbox. The sandboxing docs say sandboxing is off by default. The documented minimal example sets mode: "non-main", scope: "session" and workspaceAccess: "none" under agents.defaults.sandbox.
  2. Understand what it covers. The Gateway process always runs on the host. Only tool execution moves into the sandbox. The docs call it "not a perfect security boundary" that limits filesystem and process access when a model misbehaves.
  3. Pick a backend. Docker is the default local backend. Podman, SSH, OpenShell and Crabbox are also documented.
  4. Deny tools you do not need. Allow and deny policies apply before sandbox rules, and sandboxing does not restore a tool you denied. To hard-disable exec, use a tool policy deny.
  5. Watch the escape hatches. tools.elevated runs exec outside the sandbox. Leave it off unless a named agent needs it.
  6. Inspect the result. openclaw sandbox explain shows the effective mode, workspaces, mounts and tool policy.
  7. Give each agent its own profile. The docs describe per-agent overrides under agents.entries.*, so a support agent does not inherit what a finance agent needs.

Phase 6. Run the audit before go-live

  1. Run openclaw security audit. The docs say it checks for drift from the defaults and reports findings with a severity and, where supported, an auto-fix.
  2. Run openclaw policy to inspect and test the tool policy.
  3. Read the secrets guidance in the security docs. It covers file permissions, workspace .env files, logs and transcripts.
  4. Keep every key named dangerous* or dangerously* out of production config.
  5. Do not bind the Gateway beyond loopback until you have read the exposure runbook. The docs cover bind, firewall, Gateway auth, Tailscale and reverse proxies.

Phase 7. Run a pilot, then widen it

  1. Pick one workflow with a clear owner, such as summarizing emails, drafting replies or answering support tickets. These are all use cases AyClaw Bots lists on its own site.
  2. Run it with five to ten people for two weeks.
  3. Review the logs and transcripts weekly. Look for requests the agent should have refused.
  4. Write down every config change in your runbook.
  5. Add the second workflow only after the first has a week with no surprises.

Phase 8. Day-two operations

  1. Updates. Use the documented update guide, back up first, and run openclaw doctor after.
  2. Monitoring. Check openclaw health and openclaw gateway status on a schedule.
  3. Model changes. Treat a model switch as a change request. Behavior differs between models.
  4. Offboarding. Remove the person from allowlists and rotate any credential they could reach.
  5. Review cadence. Re-run openclaw security audit after every config change and every upgrade.

When a partner makes sense

The checklist above is a few weeks of part-time work for a good platform engineer. For regulated teams, the harder part is writing policy. AY Automate runs OpenClaw and NemoClaw enterprise setups in four phases: discovery and audit, architecture and policy, deploy and integrate, and harden and scale. Its page lists custom YAML sandbox policies, a deny-by-default posture, monitoring dashboards and SLA-backed support, with a pilot in about two weeks and a typical full engagement of four to eight weeks. The first call is free and lasts 30 minutes.

NVIDIA's NemoClaw is the other piece you will see mentioned. NVIDIA describes it as an open source reference stack for running supported agents, OpenClaw by default, inside OpenShell sandboxes. The repository calls it an alpha project, so plan for change if you adopt it.

FAQ

How long does an OpenClaw enterprise setup take?

The install itself takes minutes. The work around it takes longer, mostly access control, sandbox choices and the first pilot. AY Automate's service page quotes about two weeks to a live pilot and four to eight weeks for a typical full engagement, depending on compliance needs and the number of workflows.

Do I need Docker to run OpenClaw?

No. The install docs list installer scripts, npm, pnpm, bun and source builds alongside containers. Docker matters for sandboxing, because the sandboxing docs name it as the default local backend, with Podman, SSH, OpenShell and Crabbox as alternatives.

Can one OpenClaw instance serve several departments?

It can, but the docs advise against mixing groups that should not trust each other. They describe one trust boundary per Gateway and say mutually adversarial users need separate Gateways, credentials and ideally separate hosts or OS users.

Is the sandbox on by default?

No. The sandboxing docs state that it is off by default and must be enabled in config. Turn it on before the agent touches anything sensitive, then confirm the result with openclaw sandbox explain.

Can I skip all of this and use a managed instance?

Yes. AyClaw Bots deploys a managed OpenClaw instance after you pick a model and connect Telegram, and its FAQ says it runs unmodified OpenClaw so you can export your data and leave. Pricing is not published yet.