I've seen a recommendation to add elevated_mode: true and sandbox: false to openclaw.json to allow the gateway to manage itself properly β kill hung processes, restart services, bind to ports. Are these valid config keys in OpenClaw 2026.4.21? If so, what do they actually do and is it safe to set them on a local WSL2 install?
#Elevated Mode & Sandbox Config Keys β Valid in OpenClaw?
1 messages Β· Page 1 of 1 (latest)
Assuming openclaw --version really says 2026.4.21, no, not as written.
These are not the supported config keys:
elevated_mode: truesandbox: false
The real 2026.4.x knobs are:
agents.defaults.sandbox.mode: "off" | "non-main" | "all"tools.elevated.enabled: truetools.elevated.allowFrom.<channel>: [...]- optionally
agents.defaults.elevatedDefault: "off" | "on" | "ask" | "full"
Important bit: elevated only matters when the session is sandboxed. It is an exec-only escape hatch out of the sandbox. It does not give the gateway root/admin powers, and it does not magically let it restart services or bind privileged ports.
So on a local WSL2 install:
sandbox.mode: "off"is basically the normal host-first behavior already.
tools.elevatedis only useful if you intentionally enable sandboxing and want a break-glass host exec path.- Restarting services, killing other users' processes, or binding ports like
<1024still depends on the actual WSL/Linux user permissions (and for service management, usually whethersystemdis enabled in WSL).
So my recommendation is:
- Personal local WSL2 install: leaving sandbox off is fine if only trusted chats can reach runtime/file tools.
- Do not enable broad
tools.elevatedaccess in shared/group contexts. - If you want isolation, use
agents.defaults.sandbox.mode, notsandbox: false.
Docs:
- https://docs.openclaw.ai/gateway/configuration-reference#agentsdefaultssandbox
- https://docs.openclaw.ai/tools/elevated
- https://docs.openclaw.ai/gateway/sandboxing
- https://docs.openclaw.ai/gateway/sandbox-vs-tool-policy-vs-elevated
If you want, paste your openclaw.json snippet plus openclaw sandbox explain, and I can sanity-check the exact safe config.