What Permissions Should Your OpenClaw Agent Have? A Permissions Checklist
Published 2026-11-15 · Updated 2026-11-15 · By The Is My Agent Safe team
Short answer
An OpenClaw agent should have only the tools its job needs. OpenClaw's docs say local onboarding sets the tool profile to full when none is configured, so start there: pick a smaller profile, deny tools you don't use, make shell commands ask a human or deny them, keep elevated off, and check the result with openclaw sandbox explain.
- Tool profile and tool policy decide which tools exist. Deny always wins.
- Exec approvals decide whether shell commands run, ask a human, or are denied.
- The sandbox decides where tools run. Elevated is an exec-only way out of it.
- openclaw sandbox explain and openclaw exec-policy show print what is actually in effect.
- Re-run openclaw security audit --deep after every change.
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
OpenClaw owners who set the agent up with the defaults and now want to know what it is allowed to do.
What decides what an OpenClaw agent can do?
OpenClaw's docs name three controls:
- Sandbox decides where tools run: in a sandbox or on the host.
- Tool policy decides which tools are available.
- Elevated is an exec-only escape hatch from ordinary sandboxing.
Exec approvals sit on top. They decide whether a shell command runs, asks a human, or is denied (exec docs).
What does OpenClaw allow by default?
OpenClaw's tool policy docs say local onboarding sets tools.profile: "full" when no profile is configured. full applies no core profile filtering. The smaller profiles are:
| Profile | What it includes, per OpenClaw's docs |
|---|---|
minimal | session_status, plus gateway for maintenance only |
messaging | messaging and session tools, plus ask_user |
coding | file, shell, web, session and memory tools, plus cron and others |
The coding and messaging profiles also implicitly allow configured MCP servers. A profile selects tools. It doesn't grant Full access execution permissions, which the chat Permissions menu controls.
The permissions checklist
1. See what is in effect now
openclaw --version openclaw sandbox explain --agent main openclaw exec-policy show --agent main
sandbox explain prints the effective sandbox mode, tool allow and deny lists, and elevated gates (docs). exec-policy show explains terminal tool policies and command approvals, and never changes permissions (approvals CLI docs).
2. Pick the smallest profile that does the job
A chat assistant may need only messaging. Add single tools back with tools.alsoAllow rather than switching to full.
3. Deny what the agent doesn't use
deny always wins. If allow is non-empty, everything else is blocked (docs). OpenClaw's hardened baseline denies group:automation, group:runtime, group:fs, sessions_spawn and sessions_send, then re-enables tools per trusted agent.
4. Decide what shell commands may do
Tool policy works on names and doesn't inspect what a shell command does. If exec is allowed, denying write doesn't make shell commands read-only. For a read-only agent, OpenClaw's docs say to deny group:runtime as well as the mutating file tools. If the agent does need a shell, the exec docs list modes from deny to full. In ask mode, allowlist matches run and everything else asks a human.
5. Keep elevated off unless you need it
Elevated lets exec run outside the sandbox, and /elevated full skips exec approvals for the session. The hardened baseline sets tools.elevated.enabled: false. If you enable it, limit tools.elevated.allowFrom to named senders. The audit checks list rates a "*" entry there critical.
6. Check the sandbox mode and mounts
In "non-main" mode, only non-main sessions are sandboxed, which OpenClaw's docs call a common surprise for groups and channels. Mounts pierce the sandbox, default to read-write if you omit the mode, and mounting /var/run/docker.sock "effectively hands host control to the sandbox."
7. Use a session permission mode
OpenClaw's permission modes are read-only, guarded ("Ask me"), workspace ("Auto-review") and full ("Full access"). full has no exec reviewer.
8. One trust boundary per gateway
OpenClaw's security docs assume one trusted operator or team per gateway. For people who don't trust each other, they advise a separate gateway and credentials, ideally separate OS users or hosts.
9. Re-run the audit
Run openclaw security audit --deep. Findings that point at permissions include tools.exec.security_full_configured, security.exposure.open_groups_with_elevated and plugins.tools_reachable_permissive_policy. See our guides to exec security full and open groups with elevated.
What our free check covers
We built a free paste check. Paste your 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.
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. The full report explains every finding and lists the fixes in order. We don't review the code inside skills or MCP servers ourselves, beyond what OpenClaw's own audit reports.
Is My Agent Safe is independent and isn't affiliated with or endorsed by OpenClaw or its maintainers.
Written with AI assistance by the Is My Agent Safe team at Your Evenings Ltd.
Where next
- AI agent security checklist for OpenClaw owners
- Can OpenClaw delete your emails?
- security.preset removed: set sandbox and exec controls directly
Frequently asked questions
What tool profile does OpenClaw use by default?
OpenClaw's tool policy docs say local onboarding sets tools.profile to full when no profile is configured, including when onboarding runs again on an existing config without a profile. The other profiles are minimal, coding and messaging.
What is the difference between sandbox, tool policy and elevated in OpenClaw?
OpenClaw's docs describe three controls. The sandbox decides where tools run. Tool policy decides which tools are available. Elevated is an exec-only escape hatch from ordinary sandboxing, and it does not grant extra tools.
How do I make an OpenClaw agent read-only?
OpenClaw's docs say that for read-only agents you should deny group:runtime as well as the mutating file tools, unless sandbox filesystem policy or a separate host boundary enforces read-only. Denying write alone doesn't stop shell commands.
Can Is My Agent Safe check this for me?
We built a free paste check. Paste your openclaw security audit --deep output 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. We don't review the code inside skills or MCP servers ourselves, beyond what OpenClaw's own audit reports.
Sources
- OpenClaw docs: tool policy configuration (checked 15 Nov 2026)
- OpenClaw docs: sandbox vs tool policy vs elevated (checked 15 Nov 2026)
- OpenClaw docs: exec tool (checked 15 Nov 2026)
- OpenClaw docs: approvals and exec-policy CLI (checked 15 Nov 2026)
- OpenClaw docs: session permission modes (checked 15 Nov 2026)
- OpenClaw docs: hardened baselines (checked 15 Nov 2026)
- OpenClaw docs: security overview and trust boundary (checked 15 Nov 2026)
- OpenClaw docs: security audit checks (checked 15 Nov 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.