Question 1

Is the latency claim realistic or just a fast machine?

Accepted Answer

Read it as an order of magnitude. The README reports a median decision step of 0.6 to 1.9 seconds on its own codebase, about 2.5 seconds end to end for a click-only task, and roughly 60 seconds for a ten-step flight search, with most of that time in page animation and the text-helper model rather than the decision. Its benchmark page itself warns the samples are small and come from one machine.

Question 2

What happens if a page tries to trick it into clicking something dangerous?

Accepted Answer

Several layers, all described as mitigations rather than guarantees. A keyword denylist halts the run on payment, purchase, delete, send, transfer, authorize or login wording in English and Chinese; a cross-domain check stops after any hostname change; decisions only count against a fingerprint of the page they were made on; and a user takeover stops everything. The README admits a denylist can be worded around and that link URLs are never inspected.

Question 3

Does it need a second model just to fill in text fields?

Accepted Answer

Only for typing, and only the string. A small OpenAI-compatible model is called when a text field needs a value and writes that string, nothing else, tested with Qwen on DashScope and GLM on z.ai, with Qwen3.8-flash the fastest in their runs. Choosing the operation and the target stays a single typed-choice request to Jev, which is what keeps the loop to one call per step.

Question 4

Will it cope with a date picker?

Accepted Answer

That is the weakest area it admits to. The limits section covers DOM-only pages, with no file uploads, canvas, iframes, shadow DOM or complex keyboard widgets, and calls date pickers and similar composite widgets the most fragile part. A date goal must be in the future, because an unbookable date makes the agent spin until the action budget stops it, which at least fails safe instead of clicking randomly.

Question 5

What do I need to run it?

Accepted Answer

Ego Lite installed and running on macOS, a TYPESAFE_API_KEY, and optionally a key for any OpenAI-compatible chat endpoint if a text field needs a value. There is no npm install and no build step; the whole agent is one file, jego.js, driven by ./run.sh.

Question 6

What data gets sent on each step?

Accepted Answer

Every step sends the page URL, title, up to 6000 characters of visible text, control labels, and the current values of ordinary input fields to api.typesafe.ai, with password, file, and hidden inputs excluded. When typing, the same page context plus your goal and recent actions also goes to the TEXT_MODEL_BASE_URL you configured.

Question 7

What limits and hard budgets apply to a run?

Accepted Answer

It stops after 60 actions or 120 model calls, and after 3 consecutive no-op actions. It handles DOM pages only: no file uploads, canvas, iframes, shadow DOM, or complex keyboard widgets, date pickers are the most fragile part, and a date goal must be in the future or the agent spins until the step budget stops it.

Ego Jev Ultrafast

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community