Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Browser Nav

    1

    Picks the next click or typed field toward a UI test goal, while your agent runs it. Early and lightly tested, so use it on test sites.

    Free

    1 useSecurity scanned
    Browser Nav

    Browser Nav

    Example session with this skill installed

    Use browser-nav to sign in as admin and open the settings page on my local test app, one step at a time.

    • Read your context and instructions
    • Compiled the browser nav

    Summary

    Goal reached in

    5 calls: type the email, type the password, click Sign in, click Settings, then done. Your agent runs each step in the browser and reports pass.

    This is real output from the packaged skill, driven over a scripted, made-up login flow (a login page, a dashboard and a settings page). For this run I pointed the skill at a tiny local stand-in that just works through Email, Password, Sign in and Settings in order. It shows the call format and the step loop. It says nothing about how a real model picks steps on a real site.

    What each call returned

    step 1  exit=0  {"action": "type",  "ref": "ref_1", "p": 0.9}
    step 2  exit=0  {"action": "type",  "ref": "ref_2", "p": 0.9}
    step 3  exit=0  {"action": "click", "ref": "ref_3", "p": 0.9}
    step 4  exit=0  {"action": "click", "ref": "ref_7", "p": 0.9}
    step 5  exit=0  {"action": "done"}
    

    The same run, step by step

    StepPageElements it was shownSkill saidYour agent does
    1/loginEmail, Password, Sign in, Forgot passwordtype ref_1types the email you gave it
    2/loginsametype ref_2types the password you gave it
    3/loginsameclick ref_3clicks Sign in
    4/dashboardSettings, Reports, Sign outclick ref_7clicks Settings
    5/settingsDisplay name, Savedonechecks the page and reports pass

    Next steps

    • Pass the current url and your history on every call, or a finished goal can look unfinished.
    • If it says unsure or error, pick the step yourself.
    • If it says blocked, stop and report fail with the last page state.
    • Never let it submit payments or destructive actions without asking you first.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Hey! This is a small helper for testing multi-step UI flows, like "sign in as admin and open settings" or a role-based access check.

    You tell it your goal and hand it the page's interactive elements (role and name for each). It tells you the single next step: click this, type into that, you're done, or you're stuck. Your own agent does the clicking with its browser tool and gives the final pass or fail. At most 25 steps.

    The loop

    1. Read the page's interactive elements with your browser tool.
    2. Pipe a small JSON request (goal, url, elements, history) into the script.
    3. Act on the answer:
      • click: click that element.
      • type: focus that element and type what the goal calls for. It never invents text, and credentials come from you.
      • done: check the goal on screen and report pass.
      • blocked: stop and report fail with the last page state.
      • unsure or error: it is not confident or unavailable, so choose the step yourself.
    4. Add what you did to history and repeat.

    It does not control a browser by itself. You need an agent with a browser tool (for example Claude in Chrome or Playwright).

    Modes, and what is sent

    • Model mode (optional). With a free Codiv OpenJev API key as OPENJEV_API_KEY, or a local model set up with jev-tools-setup, a small decision model picks the step. It is sent the goal, the page URL, your last 8 actions and the elements' roles and names (up to 200), to api.codiv.ai or a loopback address for a local model. Secret-shaped strings are redacted first and the key is only ever sent to those hosts. Don't point it at pages showing private data.
    • No key. It returns unsure with the element whose name shares the most words with your goal, as a hint only, or error if nothing matches. Nothing is sent anywhere.

    What you need

    Python 3.10+. Standard library only, nothing to install.

    Honest limits

    • This is early. It was tested on one synthetic login page, five steps for "sign in as admin and open the settings page" (type email, type password, click Sign in, click Settings, then done). An earlier run said blocked on the final page because the model's goal-met score wobbled between identical calls; the rule that "nothing left to click and the goal probably met" counts as done came from that. It has never been run against a real browser session. Treat every step as a suggestion.
    • Pass the current url and recent history, or a finished goal can look unfinished.
    • Answers vary slightly between identical calls.
    • Never submit payments or take destructive actions without asking the user.
    • It appends one small line per model call (timestamp, number of questions, milliseconds, tokens, or an error type) to ~/.jev-tools/log.jsonl, or to the path in JEV_LOG. Page content and keys are never logged.

    Part of jev-tools

    One of several skills from the free, MIT-licensed jev-tools project: github.com/Rcidshacker/jev-tools

    Needs Python 3.10+ and an agent with its own browser tool. It only suggests the next step and never drives a browser itself. Max 25 steps per goal and up to 200 elements per page. Early: tested on one synthetic login page only, never on a real browser session. Answers vary slightly between identical calls. With no key it gives a keyword hint only. Tested by the author on Windows only.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    Codiv
    Python

    Permissions

    Terminal / Shell
    Browser
    Environment Variables
    Read Files
    Write Files
    Network Access

    Allowed Hosts

    api.codiv.ai

    File Scopes

    ~/.jev-tools/**

    Shell runs the bundled Python script. Browser: the skill's loop has your agent use its own browser tool to read a page's interactive elements and click or type; the script itself never touches a browser. Files: it reads the optional config and credentials files that jev-tools-setup writes in ~/.jev-tools, and appends one small log line per model call there (never page content or keys). Environment: reads OPENJEV_API_KEY (and JEV_MODE, JEV_LOG) if you set them. Network: only in optional model mode, a POST to api.codiv.ai (or a loopback address for a local model) with the goal, URL, last 8 actions and element roles and names, secret-shaped strings redacted first; redirects are refused. With no key it makes no network calls.

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    Given a goal and the page's interactive elements (role and name), it names the single next step: click an element, type into one, done, blocked, or unsure. Your own agent runs each step with its browser tool and gives the final pass or fail. It does not drive a browser by itself.

    No. Without a key it returns "unsure" with the element whose name best matches your goal, as a hint only, and sends nothing anywhere. With a free Codiv OpenJev key (OPENJEV_API_KEY) or a local model set up with jev-tools-setup, it sends the goal, the page URL, your last 8 actions and up to 200 elements' roles and names to api.codiv.ai (or a loopback address). Secret-shaped strings are redacted first, but don't point it at pages showing private data.

    It is early. It was tested on one synthetic login page: five consecutive steps for "sign in as admin and open the settings page", ending in done. An earlier run reported "blocked" on the last page because the goal-met score wobbled between identical calls, which is why "nothing left to click and the goal probably met" now counts as done. It has never been run against a real browser session, so treat each step as a suggestion and use it on test sites.

    No. It only names which field to fill. It never invents text, and credentials come from you. Your agent does the typing with its own browser tool. Never let it submit payments or take destructive actions without asking you first.

    It is a standard SKILL.md that runs a Python script, so any agent that follows the SKILL.md standard, can run Python, and has a browser tool that lists interactive elements should work. I built it for Claude Code, where the skill folder is available as ${CLAUDE_SKILL_DIR}, and I have not tested other agents.

    New in Testing & Debugging