Question 1

What does Jev decide, and can I see why it picked an agent?

Accepted Answer

The README describes one decision, which agent should handle the task, and a flow of pick, optional approval, launch, then a prompt to enter when you are done. The approval step fires when the noul probability passes 0.7, so low-confidence picks proceed without asking. Beyond that there is no explanation output documented, and no recorded reason for the choice.

Question 2

What has to be installed, and does it need accessibility access?

Accepted Answer

pip install -r requirements.txt and playwright install chromium, then the key in secrets.env as JEV_API_KEY. On macOS you also enable Safari's Develop menu setting for JavaScript from Apple Events, and grant Terminal accessibility permission if you want desktop auto-type. On Windows and Linux the Playwright profile lives at ~/.jev-router/browser-profile and you sign in once, with xdotool needed for desktop paste on Linux.

Question 3

What happens when the desktop app is not installed?

Accepted Answer

It falls back to the web version for that entry, which is why the agent table carries a web column: Claude goes to claude.ai, ChatGPT to chatgpt.com, Cursor to cursor.com and Antigravity to gemini.google.com. Those are sites rather than the agent products, so you are signing into a browser profile rather than driving the desktop client.

Question 4

Can I test the routing before it starts opening windows?

Accepted Answer

Yes. python test_prompts.py is described as a routing smoke test, and the --launch flag is the thing that also opens the agents, so the plain invocation exercises the decision path only. For the steady state the flow ends by asking you to press enter when you are done, which is the point at which control comes back to you.

Question 5

Which agents are actually in the catalog?

Accepted Answer

Four keys: claude, chatgpt, cursor and antigravity, each mapping to a desktop app and a web fallback. The catalog itself is OS-agnostic and the platform-specific pieces live in launcher/, so the same router runs on Mac, Windows and Linux without the entry list changing.

Question 6

Where does the key live, and what does one full run do?

Accepted Answer

JEV_API_KEY goes in a secrets.env file, which has to stay out of version control. Running python router.py asks Jev which agent should take the task, waits for approval when the noul probability passes 0.7, then opens the app as a new chat and sends the prompt, or the web fallback if the app is missing, and finally waits for you to press enter when you are done.

Question 7

Is there any evidence about how good the routing actually is?

Accepted Answer

No, and that is worth saying before you rely on it. The README is about fifty lines and documents the flow only, with no accuracy figures for the pick itself, so there is nothing to judge it against. test_prompts.py exercises the decision path, and adding --launch extends it to actually open the agents, but neither reports how often the choice was the right one.

Jev Router — AABBAASS1

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community