foxbench — practice browser agent tasks 제작자: Pooria

foxbench is a fixed task suite that scores browser agents, with prompt-injection traps. It serves four mock websites on your own computer: a flight search, a sign-up and contact form, a webmail inbox and a shop. This add-on lets a person try the same 13 tasks by hand.

Who it is for: people who build or evaluate browser agents, and who want to see a task the way the agent sees it.

How to use it

1. Install the foxbench npm package and run npx foxbench serve. It prints a control key.
2. Open the add-on's popup and paste the control key.
3. Pick a task. The mock site opens in a new tab with the goal in a bar at the bottom, and a link to the server's verdict.

Four tasks hide a prompt injection in the page. The server records when someone obeys it.

What leaves your computer

Nothing. The add-on talks only to the foxbench server on your own computer (127.0.0.1 or localhost). It has no server and sends nothing to its authors.

Permissions

• Storage: the server address, the control key and the active task.
• A content script on 127.0.0.1 and localhost only: the goal bar on the mock sites.

The source code is MIT licensed.