Question 1

Is it actually faster than a small LLM, or just cheaper?

Accepted Answer

Faster, on a narrow task. Over the same four page states with 24 calls each, Jev ran at a 153ms median and 181ms p90 against 678ms and 1050ms for gpt-4o-mini in JSON mode at temperature 0, and both were 24/24 correct. The author is careful that this is one machine and one simple task, and that the element-table and action overhead is identical for any model, so end-to-end gains are smaller than the per-decision gap.

Question 2

What happens when the hunch is weak?

Accepted Answer

It exits 2 with a structured report instead of guessing: a reason from a fixed set including low_confidence, risky_click, no_progress, stuck_waiting, left_allowed_origin and done_not_verified, plus the visible text, the history of actions and the element table. That is designed as the hand-off point to an LLM or a human, and the same agent can use --dry-run as a cheap second opinion on a page without touching it.

Question 3

Can it type arbitrary text, or run anything it invents?

Accepted Answer

Neither. Jev generates no text, so hunch types only the values you pass with --value NAME=env:VAR or NAME=literal:TEXT, and free-form typing needs an LLM. Model output never becomes code: Jev can only choose among refs and value names that hunch itself listed, nothing it returns is ever used as a selector, JavaScript or a shell command, and your values go straight to the browser while Jev only sees the field names.

Question 4

Is it safe to point at a browser I am logged into?

Accepted Answer

Several guards are enforced in code rather than in a prompt: success must be verified against --done-url or --done-text before a DONE is believed, the confidence floor defaults to 0.75, irreversible clicks need two independent signals to clear, an origin allow-list stops it leaving the starting origin, two actions that change nothing end the run, and field contents are reduced to filled or empty. Set --allow-risky if you want to override the click guard. There is also --no-page-text to send only the element table.

Question 5

How do I install it and what does it depend on?

Accepted Answer

Install the browser layer with npm install -g agent-browser plus agent-browser install, then hunch itself with pipx or uv tool install from the git URL. hunch is pure standard library, Python 3.10 or newer, with zero dependencies, and is MIT licensed. You also need a TypeSafe key in the environment, or a HUNCH_KEY_CMD command that prints one so it can live in a vault.

Question 6

What does a real run look like, end to end?

Accepted Answer

The demo fills in and submits a web form in 3.4 seconds: four decisions of 114 to 298 ms each, 675 ms of deciding in total at about $0.0002, then the result verified in code. The JSON output shows each step with its operation, op_conf, target ref and jev_ms, and closes with status done_verified, 3 steps, 3 Jev calls and 2.78 seconds.

Question 7

How is hunch meant to sit under a bigger agent?

Accepted Answer

As the fast path. Let it run the routine stretch of a task and take over only when it exits 2. hunch --dry-run answers what would you click here for the current page without touching it, which makes it a cheap second opinion inside an existing loop. A ready-made SKILL.md teaches Claude Code or any agent that reads skills when to reach for it and how to act on an escalation.

Hunch — HAR5HA-7663

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community