How an AI agent uses Agent safety
The part that justifies the price: a valve between a language model and your logged-in browser.
Updated · T3rnel Browser 1.3.0
What the agent can do
- An approval gate that refuses. Tools that change pages return
APPROVAL_REQUIREDnaming the exact action requested. Nothing runs until you approve, and an approval covers one action — not every future one. - Credential redaction that stays useful. Password and token-like values never reach the model, but the field's length does, so an agent can confirm a fill landed without seeing the value. Redaction that hides everything makes verification impossible, and an agent that cannot verify retries and double-submits.
- Prompt-injection flagging with coverage. Page text is scanned for instruction-override patterns and reported as suspected, along with how many characters were examined. A clean verdict over zero characters reads exactly like a passing scan, so the two are never conflated.
- Delta snapshots. After the first snapshot, only the structural difference is sent. This is where the token bill actually goes.
- Time travel. Every step keeps a compressed DOM snapshot, console and network deltas, and the exact tool arguments. When a run goes wrong at step 42, scrub to 41.
The tools
| Tool | Risk | Tier |
|---|---|---|
browser_firewall_check | low | Pro |
browser_firewall_set_policy | medium | Pro |
browser_firewall_get_policy | low | Pro |
browser_dom_compress | low | Pro |
browser_dom_delta | low | Pro |
browser_timetravel_snapshot | low | Pro |
browser_timetravel_list | low | Pro |
browser_timetravel_inspect | low | Pro |
browser_timetravel_export | low | Pro |
Calling it over MCP
{ "method": "tools/call", "params": { "name": "browser_firewall_check", "arguments": {} } }Connect a client first: add the MCP Session Bridge. The extension then exposes 97 tools against your signed-in tabs.
Questions
Which tools does Agent safety use?
browser_firewall_check, browser_firewall_set_policy, browser_firewall_get_policy, browser_dom_compress, browser_dom_delta, browser_timetravel_snapshot, browser_timetravel_list, browser_timetravel_inspect, browser_timetravel_export.
Is Agent safety free?
Partly. browser_firewall_check, browser_firewall_set_policy, browser_firewall_get_policy, browser_dom_compress, browser_dom_delta, browser_timetravel_snapshot, browser_timetravel_list, browser_timetravel_inspect, browser_timetravel_export need Pro ($29.99 once); none of the others work on the free tier.
Does the agent have to ask before using Agent safety?
Each tool has a risk level: browser_firewall_check is low; browser_firewall_set_policy is medium; browser_firewall_get_policy is low; browser_dom_compress is low; browser_dom_delta is low; browser_timetravel_snapshot is low; browser_timetravel_list is low; browser_timetravel_inspect is low; browser_timetravel_export is low. The approval gate is off by default and can be armed in Settings.