Running OpenClaw for Multiple Clients: Separate Gateways, Profiles and Hosts
Published 2026-10-20 · Updated 2026-10-20 · By The Is My Agent Safe team
Short answer
Give each client their own Gateway. OpenClaw supports one trust boundary per Gateway and says mutually untrusted users need separate Gateways, ideally on separate OS users or hosts. Several agents in one Gateway are for people who trust each other. On a shared host, prefer a separate OS user per client. If clients share one OS user, give each a named profile and base port at least 120 apart.
- One Gateway is one trust boundary. Agents inside it share session access by default, so don't put two clients on one Gateway.
- Strongest split: one host or VPS per client. Next: one OS user per client. Lightest: one named profile and port per client under one OS user.
- Keep config path, state directory, workspace and ports unique per Gateway, and leave at least 120 ports between base ports.
- Give each client their own channel accounts, model keys and spend limits, and audit each profile separately.
Paste your audit: run openclaw security audit --deep and paste the 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, freelancers and MSPs who set up or run OpenClaw for more than one client, and want to know how far apart those installs need to be. If you only run your own agent, the 12-point checklist is the better start.
The rule: one trust boundary per Gateway
OpenClaw's security trust model sets the scope:
- Supported: one trust boundary per Gateway, meaning a single operator or a team whose members trust each other. OpenClaw prefers one OS user, host or VPS per boundary.
- Not supported: one shared Gateway or agent used by mutually untrusted or adversarial users.
- For separation between such users: separate Gateways, and ideally separate OS users or hosts.
We treat two separate client businesses as mutually untrusted, so give each client its own trust boundary.
Why one Gateway with several agents isn't a client boundary
OpenClaw can run several agents in one Gateway, each with its own workspace, state directory, auth profiles and session store. That's useful for personas inside one business. It doesn't separate clients, for reasons the docs spell out:
- Session tools reach across the whole Gateway by default. Per the trust model,
tools.sessions.visibilitydefaults toallandtools.agentToAgent.enableddefaults totrue. An unsandboxed, tool-enabled agent can list, read, search and message every agent's sessions, including other users' transcripts. You can narrow this, but the trust model still says strong isolation requires separate Gateways. - A workspace isn't a hard sandbox. The multi-agent docs say each agent's workspace is the default working directory. Absolute paths can reach other locations on the host unless sandboxing is enabled.
- Operator access isn't per-client. Inside one Gateway, authenticated operator access is a trusted control-plane role, not a per-user tenant role. Named operator roles are collaboration guardrails, not tenant isolation.
- Everyone who can message a tool-enabled agent shares its tool authority. That's fine for teammates who trust each other. It's why separate clients can't share an agent.
- A session key routes messages. It doesn't authorize anyone.
Three ways to split clients
From strongest to lightest:
| Setup | What's separate | When it fits |
|---|---|---|
| One host or VPS per client | Everything: OS, files, network, credentials | Clients with sensitive data (mail, CRM, finance), or any client who asks for it. OpenClaw's trust model prefers one OS user, host or VPS per trust boundary. |
| One OS user per client on a shared host | Files and processes, through normal OS permissions | Several small clients on one server you run well. Each user runs its own Gateway. |
| One named profile per client under one OS user | Config, state, workspace, service name and ports | Lower-risk clients, test installs, or a rescue bot. All profiles still run as the same OS user. |
The third option keeps OpenClaw's own files apart, and that's all. OpenClaw's trust model says to "prefer one OS user/host/VPS per boundary", and its multi-agent docs say each agent's workspace is "the default cwd, not a hard sandbox". So don't rely on separate profiles under one OS user to keep clients' files apart if an agent has shell or file tools without a sandbox.
Running several Gateways on one host
OpenClaw's multiple gateways page covers the mechanics. Give each extra Gateway its own named profile and base port:
openclaw --profile clienta setup openclaw --profile clienta gateway install --port 19789 openclaw --profile clientb setup openclaw --profile clientb gateway install --port 19909
The global --profile <name> flag keeps state under ~/.openclaw-<name> (CLI reference). Per the multiple-gateways page:
| Keep unique per Gateway | Why |
|---|---|
OPENCLAW_CONFIG_PATH | Its own config file |
OPENCLAW_STATE_DIR | Its own sessions, credentials and caches |
agents.defaults.workspace | Its own workspace root |
gateway.port (or --port) | Its own base port |
| Derived sandbox, browser and CDP ports | Each instance uses ports up to base + 110 |
Three details that catch people out:
- Leave at least 120 ports between base ports, so the derived sandbox, browser and CDP ports never collide.
- A separate state directory alone doesn't separate a managed service. Service names follow the profile, not the state directory. Use a named profile and unique ports.
- Don't pin
browser.cdpUrlto the same value on several instances. Each one needs its own browser control port and CDP range.
To confirm what's running, use openclaw --profile <name> gateway status --deep and openclaw --profile <name> gateway probe. A multiple reachable gateway identities detected warning is expected when you deliberately run several isolated Gateways.
Per-client credentials and channels
Separate Gateways only help if the credentials are separate too:
- Channel accounts: each client gets their own bots and numbers. OpenClaw's rescue-bot example gives the second Gateway its own Telegram bot token.
- Model keys and spend limits: one key per client, ideally in the client's own provider account, each with its own spend limit. See giving agents API keys without exposing them.
- Remote access: each Gateway stays on loopback and you reach it over an SSH tunnel or Tailscale Serve, per the exposure runbook. Don't open a public port per client.
Never reuse one key, bot or SSH key across clients. One leak would then become an incident for every one of them. Our incident runbook covers what to do if that happens.
Audit and update each client separately
Each Gateway has its own config, so audit each one on its own. Profiles that run from the same OpenClaw install report the same openclaw --version:
openclaw --profile clienta security audit --deep --json > clienta-audit.json openclaw --version
Keep every client on the latest stable OpenClaw release, 2026.10.5 (October 18, 2026). Any client below 2026.8.1 is missing fixes for advisories rated high. Our update guide has the backup and dry-run steps, and weekly checks for clients turns this into a routine.
What our free check covers
The free check at /check is paste-only and works one client at a time. Paste one profile's openclaw security audit --deep output and openclaw --version. It explains the findings, checks for risky settings such as API keys handed to skills, and checks the version against published security advisories. Nothing is installed into the agent you're checking. Green means no issues found in the checks we ran. See Limits of the check.
It doesn't check how your clients are separated from each other, doesn't detect malicious skills, and doesn't review MCP servers or skill code. The check is free, and a full report is $29.
You can buy a founding seat on /agencies. $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. Offer conditions.
Where next
- OpenClaw: Security trust model, Multiple gateways, Multi-agent routing
- OpenClaw security for agencies and MSPs
- How to audit a client's OpenClaw
- Client's OpenClaw compromised? Incident runbook
Frequently asked questions
Can several clients share one OpenClaw Gateway?
Not as separate clients. OpenClaw's trust model supports one trust boundary per Gateway: a single operator or a team whose members trust each other. It says one shared Gateway or agent used by mutually untrusted users isn't supported, and that adversarial-user isolation needs separate Gateways, ideally on separate OS users or hosts.
Aren't OpenClaw's multiple agents isolated from each other?
Each agent has its own workspace, state directory, auth profiles and session store. But OpenClaw's trust model says session tools reach across the whole Gateway by default (tools.sessions.visibility is all and tools.agentToAgent.enabled is true), and the multi-agent docs say a workspace is the default working directory, not a hard sandbox. Use separate Gateways for strict separation.
How do I run more than one Gateway on the same server?
Give each one a named profile and its own base port, for example openclaw --profile clienta setup and openclaw --profile clienta gateway --port 19789. OpenClaw's multiple-gateways page says to keep the config path, state directory, workspace and ports unique per instance, and to leave at least 120 ports between base ports.
Is a separate profile enough to keep clients apart?
A profile separates config, state, workspace, service name and ports. All profiles still run as the same OS user, and OpenClaw's trust model prefers one OS user, host or VPS per trust boundary. Use separate OS users or hosts when clients are separate businesses with sensitive data.
Can I check every client with Is My Agent Safe?
Yes, one at a time. Run openclaw --profile <client> security audit --deep and paste the output, with openclaw --version, into the free check at /check. It explains the findings, checks for risky settings and checks the version against published security advisories. It doesn't detect malicious skills or review MCP servers or skill code.
Sources
- OpenClaw docs: security trust model (one trust boundary per Gateway)
- OpenClaw docs: security overview
- OpenClaw docs: multiple gateways (isolation, ports and profiles)
- OpenClaw docs: multi-agent routing
- OpenClaw docs: CLI reference (global --profile flag)
- OpenClaw docs: Gateway exposure runbook
- OpenClaw docs: openclaw security (CLI)
- OpenClaw release v2026.10.5 (GitHub, Oct 18, 2026)
- OpenClaw release v2026.8.1 (Aug 31, 2026)
Get the plain-English agent security brief
Agents & AI Security, explained: a short, plain-English email when we have something useful. Free. Unsubscribe any time.
We'll only use your email to send the newsletter. The newsletter is sent by beehiiv, which records opens and link clicks. Privacy policy.