OpenClaw (lobster) execution commands are always subject to approval, indicating that the host, security or ask policy of the exec tool has been tightened. This isn't necessarily a bad thing: approvals prevent agents from directly executing high-risk commands. What you need to do is to adjust the policy based on task risk, rather than turning off approvals altogether.
Let's look at the current strategy first
openclaw config get tools.exec.hostopenclaw config get tools.exec.securityopenclaw config get tools.exec.askopenclaw gateway restartand then take the test again
How to adjust it more appropriately
| Goal | suggested |
|---|---|
| Just want to fix the execution locally | Set tools.exec.host=gateway |
| Common security commands are exempt from approval | Use security=allowlist to explicit allowlist |
| Unfamiliar commands require confirmation | Use ask=on-miss |
| The test environment is fully automated | Consider security=full + ask=off only on isolated machines |
If the log appears SYSTEM_RUN_DENIED: approval required, go to the approval request to confirm the command content. Don't turn your production computer into an unlimited-privilege environment with a single click, especially if the agent can receive external messages.
Official open source address: https://github.com/openclaw/openclaw.
In actual use, you can put the "Read Status" command into the allowlist, such as viewing logs, column directories, and running tests. Keep the "write, delete, upload, send" command for approval. This will neither affect efficiency nor release the risk at once.
