Choose the next browser action

Editorial kit referencing Browser Use's public Jev Ultrafast repository. Jev selects among allowed actions; browser automation still belongs to the host agent and must stay permissioned. The repository was source-screened, not code-audited or executed here.

Read source

What you'll make

Sketch a typed action menu for a browser agent, with human approval on high-impact moves.

A constrained Choice schema and a checklist. Not a copy of Browser Use internals.

Jev decides

Selects among actions you enumerated.

Your code acts

The host agent/browser tool executes only after your code allows it.

Agent prompt

Give your agent a useful brief.

Key-safe by design

Codex prompt

You are helping a developer prototype a Jev-powered application.

Goal: Design a typed browser-action selector with human approval for high-impact actions.

Product truth:
- Jev (TypeSafe System One) returns typed decisions and calibrated probabilities. It does not write emails, generate UI copy, play games by emitting pixels, or call Zapier.
- Your job is to write application code. Jev only answers questions whose possible answers are defined in advance.
- Use the official JS SDK documented at https://github.com/typesafe-ai/typesafe-sdk-js. Read the current README before inventing method names.
- Keep API keys in server-only environment variables. Never put TYPESAFE_API_KEY in client code, screenshots, git, or logs.
- Ship a working mock adapter first. Live TypeSafe calls are optional and must be clearly labeled.

Required behavior:
1. Inspect the current project and only change what is needed.
2. Define a small typed Choice or Score schema. Include an explicit uncertain / human-review path.
3. Implement mock mode with fixtures that include ordinary, ambiguous, and conflicting examples.
4. Add a server-only adapter that calls TypeSafe only when TYPESAFE_API_KEY is present.
5. Application code must validate the typed response and decide the action. Failed provider calls must not trigger side effects.
6. Do not send, delete, archive, or mutate real user accounts in this first version.
7. Add tests for the schema and mock path. Report what was run and what remains untested.

Official source to read: https://github.com/typesafe-ai/typesafe-sdk-js

Keep the action enum small and testable.

When finished, write a short Verification section distinguishing mock results from any live API call.

Copies instructions into your clipboard. It does not run or install anything.

Setup and guardrails

Start small. Keep the boundary clear.

Prerequisites

Read the referenced repository's README
Do not grant unbounded browser control

Setup steps

Enumerate allowed actions with no catch-all 'do anything' option.
Require confirmation for purchase, delete, submit, or login.
Log state, choice, and whether the human overrode it.

Mock mode

Choose from a static list of demo actions without touching the network.

Live mode

Only after you independently review the community repo and your own permissions model.