JevOnly AI Agent

JevOnly

Pure Jev that can "type" and drive towards task completion. No LLM anywhere: code builds every option from the page and the goal, Jev only picks — and that is enough to fill forms, type, and finish the task.

git clone https://github.com/buluoray/JevOnly.git && cd JevOnly && ./run.sh

Needs Python 3.12+, Node.js 22+ and a TypeSafe API key (tsk_…). run.sh creates a venv, installs Chromium, and opens the live viewer at http://127.0.0.1:7791: paste the key, a start URL and a goal, press Run, and watch every observation, vote, action, verification, copy and undo as it happens.

Without the viewer:

export TYPESAFE_API_KEY=tsk_…
jevonly run --start https://en.wikipedia.org \
  --goal "On Wikipedia, find the tallest building in Seattle and the tallest building in Portland, Oregon. Note the height in meters of each." \
  --variant noaccept_kb --out events.jsonl

That goal ends in about 20 steps, 80 Jev calls and 30 seconds with the answer 286; 166.

Demo

The recordings above, in full quality (click to play):

https://github.com/user-attachments/assets/4f21c5ab-c840-46dc-9337-0de8c4a16df6

https://github.com/user-attachments/assets/11ad01fc-4a23-4091-b3d6-d8129c88c694

How it works

No planner model, no helper LLM, no free text. Every step is a closed question to Jev over options the code enumerated from the page, the goal and a register of copied values:

1. Observe the page (accessibility snapshot of enabled controls and visible text).
2. Vote: one request answers done?, off path? and which action next?
3. Bind values: a field can only receive a fact, a span of the goal (a name, a date, or a stretch of the goal's own words Jev narrows down to), or a value copied from the page — never invented text.
4. Act and verify: Chromium performs the action; Jev compares before and after. Failures are undone.
5. Stop when Jev says so, when every value the goal asked for is in hand, or when a code-owned check passes.

Irreversible actions (a risk vote ≥ 0.5) go through a gate: refuse (CLI default), ask (viewer default), or allow. Details: How it works · Architecture · Events · FAQ.

The rules in jevonly.core are environment-agnostic; the browser (jevonly.envs.browser) is the first environment. See Writing an environment.

QA regression suites

The same loop, pointed at a site you own, is a cheap regression tester: a suite of closed tasks, each run several times, one pass/fail row per case, every run's event log kept for replay.

jevonly qa suite.jsonl --repeat 3 --out qa-out/

One case per line:

{"id": "cart-badge", "start": "https://staging.shop.example/", "facts": {"sku": "A-1"},
 "goal": "Add the product with the given sku to the cart and stop when the cart badge shows 1",
 "expect": {"text": "Cart (1)", "answer": ["1"]}}

expect.text / expect.url_contains / expect.selected is a code-owned check on the page — the run passes only when the page really shows it, whatever Jev says. expect.selected names a control (by accessible name) whose ARIA state must read selected, pressed or checked, for end states that are a choice rather than text. expect.answer lists substrings the reported answer must contain. The command prints a table (pass count, median steps, seconds, Jev calls, tokens, estimated cost, why a run failed), writes qa-report.json + qa-report.md, and exits non-zero when any run failed. A case costs about a cent per run (measured: $0.014 for a 20-step task).

Safety and privacy

• The viewer binds to 127.0.0.1 only; the key stays in the tab's sessionStorage and the local process.
• The goal, facts and page observations are sent to TypeSafe. Point it at pages you are willing to send there.
• Nothing irreversible runs by default (see the gate above). Screenshots and replays never leave the machine.

Limits

• English-centric goal splitting; sites that block automation are treated as walls, not bypassed.
• One run at a time per viewer; without a code-owned check, stopping is Jev's judgment.

Development

make install && make test && make lint

Apache-2.0. Docs start at docs/README.md.