NewOpenHuman
Agent platform that uses Jev to choose tools and browser actions.
Added to Jevfast
NewHow it uses Jev
OpenHuman routes browser-control decisions through its TinyHumans Jev evaluator. For each step, TinyBrowser sends the user's goal, a fresh page snapshot, and recent action outcomes to Jev. Jev chooses a typed browser operation and separately judges whether the goal is visibly complete; TinyBrowser executes the selected operation and independently checks the completion signal.
What Jev decides
Choose a browser action and verify visible completionExample answers · not a recorded Jev response · Source ↗
Question 1 · operation
YOUR APP
INSTRUCTION
Choose the next allowed browser operation from the current observed page and user goal.
STATE
The request state contains goal (string), page (object with url, title, accessibility_tree, truncated), recent_actions (array of up to 8 records with step, operation, target, input, page_changed), done_unconfirmed (boolean), and fill_already_entered (boolean).
JEV · CHOICE
- CLICK
- FILL
- CHECK
- SCROLL_DOWN
- SCROLL_UP
- BACK
- WAIT
- BLOCKED
- DONE
Question 2 · goal done
YOUR APP
INSTRUCTION
Judge whether the visible page proves all requested requirements are complete.
STATE
The request state contains goal (string), page (object with url, title, accessibility_tree, truncated), recent_actions (array of up to 8 records with step, operation, target, input, page_changed), done_unconfirmed (boolean), and fill_already_entered (boolean).
JEV · NOUL
YesNo
App workflow
- Describe the browser task
A user gives the browser agent a goal. The host routes the task through the configured provider; the TinyHumans integration uses Jev for its typed decisions.
- Observe and choose
TinyBrowser sends the current URL, title, accessibility tree, and recent actions. Jev returns an operation and, when needed, a target or supplied input selection.
- Execute and verify
The controller performs the selected browser action, repeats within its configured step limit, and checks visible completion. Consequential actions can require confirmation.
Why it is interesting
It makes Jev the decision layer of an observable, multi-step computer-use agent rather than a chat interface. The controller limits the run, treats visible completion as a separate judgment, and gates consequential actions, while the browser adapter maps typed choices to concrete controls.