OpenClaw Client Onboarding: The Access Decisions to Agree Before You Install
Published 2026-10-20 · Updated 2026-10-20 · By The Is My Agent Safe team
Short answer
Agree the access decisions with the client before you install and write them down. Decide who can message the agent, whether DMs share one session, which tools it gets, whether tools run in a sandbox, which accounts it uses and how you reach the Gateway. OpenClaw's local onboarding sets the full tool profile when none is set, so choose deliberately.
- Who can message it: DM policy (pairing is the default), group allowlists and mention gates.
- Whose conversations stay apart: session.dmScope, which defaults to one shared main session.
- What it can do: tool profile, exec and elevated settings, and sandboxing, which is off by default.
- Where it runs and who can reach it, which accounts it uses, and what your contract with the client says about their data.
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 consultants about to install OpenClaw for a new client. This page covers the decisions to agree before the install. For checking an install you're about to hand over, use How to audit a client's OpenClaw. For splitting several clients, see Running OpenClaw for multiple clients.
Most of these choices are settings in OpenClaw's config. Agreeing them up front means the client knows what the agent can do, and you have a written record to check the install against later.
1. Who can message the agent?
OpenClaw's access control docs list four DM policies for every DM-capable channel:
dmPolicy | What it does |
|---|---|
pairing | The default. Unknown senders get a pairing code, and the bot ignores them until approved. |
allowlist | Unknown senders are blocked, with no pairing step. |
open | Anyone can DM the bot. Needs "*" in the channel allowlist as an explicit opt-in. |
disabled | Inbound DMs are ignored. |
Agree with the client:
- Which people can DM the agent, by name. The docs recommend pairing plus allowlists for DMs.
- Which groups it joins, and whether it needs a mention to reply. The docs say to keep sender allowlists or mention gating in rooms where strangers can join or post.
- Who can use slash commands. The docs say that if a channel allowlist is empty or includes
"*", commands are effectively open for that channel.
2. Whose conversations stay apart?
By default, OpenClaw routes all direct messages across channels into one main session. If more than one person at the client will message the agent, the docs say to set session.dmScope: "per-channel-peer", so each sender gets their own DM context.
The same page says this is a messaging-context boundary, not a host-admin boundary. If people who share the Gateway don't trust each other, the docs say to run separate Gateways, one per trust boundary. That's covered in Running OpenClaw for multiple clients.
Also agree what context reaches the model. contextVisibility defaults to all, which keeps quoted text, thread history and forwarded content as received. allowlist filters it to senders who pass the allowlist checks.
3. What can the agent do?
This is the decision with the most impact, and the one most often left at the default.
- Tool profile. OpenClaw's tool policy docs say local onboarding sets
tools.profile: "full"when no profile is configured. The other built-in profiles areminimal,codingandmessaging. Pick the smallest one that does the client's job, then add tools one by one. - Shell and elevated tools. OpenClaw's hardened baseline denies exec (
exec: { security: "deny", ask: "always" }), turns off elevated tools and limits file tools to the workspace. Agree which of these the client's work really needs. - Sandboxing. OpenClaw's sandboxing docs say it's off by default. When on, tool execution moves into a sandbox backend and the Gateway stays on the host. The docs call it "not a perfect security boundary", but say it materially limits filesystem and process access.
- Accounts the agent acts as. Which mailboxes, calendars, CRMs and chat accounts it gets. For phone-number channels, the hardened baseline suggests a separate number from anyone's personal one.
OpenClaw's running the audit page puts "anything open plus tools enabled" first in its fix order. Open messaging with broad tools is the combination to avoid.
4. Where does it run, and who can reach it?
Agree the host and the access pattern before you install:
- Host: the client's machine, their VPS, or one you run for them. OpenClaw's trust model supports one trust boundary per Gateway and prefers one OS user, host or VPS per boundary.
- Access: OpenClaw's exposure runbook recommends keeping the Gateway on loopback, reached over an SSH tunnel or Tailscale Serve, rather than a public port.
- Version: install the latest stable release, 2026.10.5 (October 18, 2026). Versions below 2026.8.1 are missing fixes for advisories rated high.
- Who holds the credentials: the client's own provider accounts and keys, or yours.
5. What does your contract say about the client's data?
If the agent will read the client's email, files or customer records, you may be handling personal data on their behalf. If you handle client data, check your contract and any data-protection rules that apply; this guide isn't legal advice. Writing down the access decisions above also gives you a plain description of what the agent can reach.
6. Write it down
Keep a one-page onboarding record per client, and check the install against it at handoff. The rows below are a made-up example, not a real client:
| Decision | Made-up example |
|---|---|
| Who can DM | Two named staff, dmPolicy: "allowlist" |
| Groups | One team channel, mention required |
| DM sessions | session.dmScope: "per-channel-peer" |
| Tool profile | messaging, plus calendar read |
| Exec and elevated | Exec denied, elevated off |
| Sandbox | On for all agents |
| Host and access | Client's VPS, loopback, SSH tunnel |
| Version | 2026.10.5 |
| Keys | Client-owned model key with a spend limit |
| Client sign-off | Name, date |
Then run openclaw security audit --deep after the install and keep the dated output with the record.
What our free check covers
The free check at /check is paste-only. Paste the client'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 your onboarding record or contract, 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: Access control, Hardened baselines, Tool policy
- How to audit a client's OpenClaw
- Running OpenClaw for multiple clients
- OpenClaw security for agencies and MSPs
Frequently asked questions
What tool profile does OpenClaw set during onboarding?
OpenClaw's tool policy docs say local onboarding sets tools.profile: "full" when no profile is configured, including when onboarding runs again on an existing config without a profile. The other built-in profiles are minimal, coding and messaging.
Who can message a new OpenClaw agent by default?
OpenClaw's access control docs say pairing is the default DM policy: unknown senders get a pairing code and the bot ignores them until you approve it. open lets anyone DM the bot and needs an explicit "*" in the channel allowlist.
Do all DMs share one conversation?
By default, yes. OpenClaw's access control docs say the main default routes all direct messages across channels into one session. For shared or multi-user inboxes, the docs say to set session.dmScope: "per-channel-peer".
Is sandboxing on by default?
No. OpenClaw's sandboxing docs say sandboxing is off by default and is controlled by agents.defaults.sandbox or per-agent settings. The Gateway process stays on the host; only tool execution moves into the sandbox.
Can Is My Agent Safe check a client install after onboarding?
Yes, one paste at a time. Paste the client's openclaw security audit --deep output and 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: access control and allowlists
- OpenClaw docs: hardened baselines
- OpenClaw docs: tool policy (profiles, allow and deny)
- OpenClaw docs: sandboxing
- OpenClaw docs: running the security audit
- OpenClaw docs: security trust model
- OpenClaw docs: Gateway exposure runbook
- 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.