You Control What Your Agent Can Do

ChatGPT can draft an email. But can it send it? Check your calendar? Organize your files? No. It's a brain in a jar — smart, but completely unable to act.

Your OpenClaw agent can actually do things. Send messages, manage files, run scripts, check your calendar. Which raises the obvious question: how do you keep it in check?

How permissions work

Like app permissions on your phone. Your agent asks before doing anything that affects your world. You see the exact command, in plain text, and approve or deny it.

Three levels of trust

• Always allowed — reading files, searching the web, checking your calendar. Low-risk, no approval needed.
• Asks first — running commands, sending messages, modifying files. The agent pauses, shows you what it wants to do, and waits. Approve once, approve always, or deny.
• Hard no — certain folders, services, or actions are completely off-limits. Not suggestions — walls. The agent can't even attempt them.

The system grows with your trust. Start by approving everything manually. Over time, allow routine tasks automatically. Keep the guardrails for anything unusual.

Compare this to the alternatives

ChatGPT, Claude, and Gemini solve safety by not letting AI do anything. No file access. No emails. No real actions. Safe because it's powerless.

That's like preventing car accidents by not building cars. OpenClaw gives you the car — with a steering wheel, brakes, and a speed limiter. You're in the driver's seat.

What it looks like in practice

You ask your agent to clean up your downloads folder:

• It scans the folder (reading — no approval needed)
• Groups 47 files by type
• Shows you the plan: "Create folders for Documents, Images, Archives. Move everything. Here's what goes where."
• You say "leave the PDFs" — it adjusts and asks again
• You approve. Done in seconds.

Feels like working with a thoughtful colleague, not fighting with software.

Change the rules anytime

Nothing is permanent. Cautious about a new integration? Require approval for everything. Trust a tool completely? Let the agent handle it automatically.

Power without control is dangerous. Control without power is useless. OpenClaw gives you both.