OpenClaw Security Checklist for Client Deployments: 10 Checks Before Go-Live

Published 2026-10-30 · Updated 2026-10-30 · By The Is My Agent Safe team

Short answer

Before a client's OpenClaw goes live, give each client its own Gateway, keep the Gateway on loopback unless there's a documented reason, run openclaw doctor, openclaw security audit --deep and openclaw health and clear every critical finding, put DMs on pairing or an allowlist, sandbox non-main sessions, deny or approval-gate host commands, check the version against OpenClaw's advisories, test from outside the network, verify a backup, and record the result with the client's sign-off.

  • One client, one Gateway, ideally its own OS user or host.
  • Loopback plus SSH tunnel or Tailscale Serve unless there's a documented reason to widen it.
  • Baseline: openclaw doctor, security audit, security audit --deep, health. Clear criticals first.
  • Channels on pairing or allowlists, groups mention-gated, non-main sessions sandboxed, host exec denied or set to ask.
  • Version checked against advisories, outside test done, backup verified, sign-off recorded.

Run the free check: after step 3, paste your openclaw security audit --deep output into our free check for a red/amber/green result with the most important finding first (Limits of the check).

Who this is for

Agencies and consultants who have just built an OpenClaw deployment for a client and are about to switch it on. It's the last check before the client's staff or customers start messaging the agent.

It sits between two of our other guides. OpenClaw client onboarding covers the access decisions to agree before you install. How to audit a client's OpenClaw covers taking over an install someone else built. This page is for the install you built, on the day it goes live.

Most of it comes from the review checklist in OpenClaw's own Gateway exposure runbook. We've put it in go-live order and added what to record for the client.

1. One client, one Gateway

OpenClaw's trust model supports one trust boundary per Gateway: one operator, or a team whose members trust each other. It says isolation between users who don't trust each other needs separate Gateways, ideally on separate OS users or hosts.

For agencies, that means each client gets its own Gateway, credentials and state directory. Don't add a second client to an existing Gateway to save a server. See Running OpenClaw for multiple clients for the layouts.

Record: host, OS user, state directory (default ~/.openclaw).

2. Pick the narrowest way in, and write it down

OpenClaw's runbook says to prefer the narrowest pattern that fits the workflow:

The runbook says to avoid direct public port-forwarding to the Gateway. Check what you set:

openclaw config get gateway.bind

Record: the pattern you chose, the port (18789 by default) and the auth mode.

3. Run the baseline and clear criticals

openclaw doctor
openclaw security audit
openclaw security audit --deep
openclaw health

These are the runbook's baseline checks. Resolve critical findings first. The runbook says to accept warnings only when they're intentional and documented for the deployment. In OpenClaw, document them with security.audit.suppressions and a reason, so they stay visible in the audit JSON under suppressedFindings.

Keep a dated copy for the client file:

openclaw security audit --deep --json > "golive-audit-$(date +%Y%m%d).json"

4. Channels: who can message the agent

The runbook treats messaging channels as untrusted input. Before go-live:

Record: every enabled channel, and whether it takes DMs, groups or webhooks.

5. Tools: what the agent can do

From the runbook's review checklist and its starting config:

The runbook's advice is to widen one control at a time. Turn on one tool or channel, re-run the audit, then the next.

6. Keys and credentials

7. Version against advisories

openclaw --version

Compare it with the current stable release on OpenClaw's releases page on go-live day, and read OpenClaw's security advisories for that version. Versions before 2026.8.1 are missing fixes for advisories rated high. Don't hand over a new deployment on an old version.

8. Test from outside the network

The settings say what you meant. An outside test shows what's actually reachable. From a network you don't control, such as a phone on mobile data, try http://<public-IP>:18789. If any page loads, the Gateway is reachable from the internet. Only test machines you own or are authorized to test. Our guide Is your AI agent exposed? has the full method.

9. Backup and rollback, tested

openclaw backup create --verify --output <private-path>

OpenClaw's backup docs say --verify validates the archive right after writing it. The runbook's review checklist also asks for rollback steps that are tested and documented. Store the archive where only you and the client can reach it, because it holds config and credentials.

10. Sign-off

One short record, agreed with the client before go-live:

CheckResult (made-up example)
Gateway per clientOwn host, own OS user
Way inLoopback plus Tailscale Serve, token auth
Deep audit0 critical, 1 warn (accepted, reason recorded)
ChannelsTelegram DMs on pairing, groups mention-gated
ToolsNon-main sandboxed, exec set to ask, elevated off
Secrets auditExit code 0
VersionChecked against advisories on go-live day: none apply
Outside testNo page loads on port 18789 from mobile data
BackupCreated and verified, rollback tested
Next reviewDate agreed, and after any update, new channel or new tool

The rows are a made-up example, not a real client. The record shows what you checked on the day. It isn't a promise that nothing can go wrong, and it goes out of date with every change. For how to present it, see AI agent risk report for clients.

What our free check covers

We built a free paste check for step 3. Paste the client's openclaw security audit --deep output and openclaw --version. It puts the most important finding first, in plain English, with how to fix that one when we have steps for it. It also looks for risky settings, such as API keys handed to skills, and checks the version against published OpenClaw security advisories. Nothing is installed into the agent you're checking.

We check the text you paste. Your browser hides many common key and token formats before sending, but not all of them, so remove any other secrets first. Green means no issues found in the checks we ran. See Limits of the check. It doesn't test from outside your network, check backups or run openclaw secrets audit. We don't review the code inside skills or MCP servers ourselves, beyond what OpenClaw's own audit reports.

The check is free, and a full report is $29. One-off $29, not a subscription. Refunded automatically if the check can't finish, or in full if you ask within 14 days. See Limits of the check.

Is My Agent Safe is independent and isn't affiliated with or endorsed by OpenClaw or its maintainers.

For agencies with several clients

We're holding 20 founding seats for agencies. The seat gives you a bulk check of up to 25 client installs (paste each install's OpenClaw audit output), re-checks of your clients against new OpenClaw advisories without pasting again, for 90 days from each install's last paste, and a PDF report for each client with your agency's name. 14-day refund on your first payment. It's for businesses located in the United States or the United Kingdom, bought by card. $99/month, price locked for your first 12 paid months. After that, it moves to the then-current Agency price, with at least 30 days' email notice. Cancel any time. It renews automatically every month until you cancel, and you can cancel online from your dashboard. Offer conditions.

See what the seat includes: Is My Agent Safe for agencies.

Where next

Frequently asked questions

Can several clients share one OpenClaw Gateway?

Not if they don't trust each other. OpenClaw's trust model supports one trust boundary per Gateway and says adversarial-user isolation needs separate Gateways, ideally on separate OS users or hosts.

Which commands should I run before a client deployment goes live?

OpenClaw's exposure runbook lists openclaw doctor, openclaw security audit, openclaw security audit --deep and openclaw health. It says to resolve critical findings first and accept warnings only when they're intentional and documented.

Is it fine to expose the Gateway port so the client can reach the dashboard?

OpenClaw's runbook says to avoid direct public port-forwarding to the Gateway. For admin access it recommends loopback plus an SSH tunnel or Tailscale Serve. If public access is truly needed, put an identity-aware proxy in front and make it the only network path.

How is this different from a handoff audit?

This checklist is for a deployment you built, before the client starts using it. If you're taking over an install someone else set up, use our handoff audit guide, which starts with an inventory before any change.

Can Is My Agent Safe run this checklist for me?

Part of it. We built a free paste check: paste openclaw security audit --deep and openclaw --version, and it puts the most important finding first, in plain English, with how to fix that one when we have steps for it, and checks the version against published advisories. It doesn't test from outside your network, check backups or review skill code.

Sources

  1. OpenClaw docs: Gateway exposure runbook (baseline checks, review checklist)
  2. OpenClaw docs: security trust model
  3. OpenClaw docs: hardened baselines
  4. OpenClaw docs: security audit checks
  5. OpenClaw docs: openclaw security (CLI)
  6. OpenClaw docs: openclaw secrets (CLI)
  7. OpenClaw docs: openclaw backup (CLI)
  8. OpenClaw GitHub security advisories
  9. OpenClaw releases (GitHub)