Skip to content

Browser Automation Options ​

AgentX's browser automation has two independent choices that people often mix up. Understanding them separately makes setup obvious:

  1. Backend — where the browser runs (the hands): System WebView vs Cloud Tunnel
  2. Decision engine — who decides what to click (the brain): Jev vs Drex vs main model

You can combine any backend with any decision engine.


Part 1: Browser backends (the hands) ​

  • What: Android's built-in System WebView, driven on-device through its remote-debugging socket. No downloads, no extra processes, 0 MB added to the app.
  • Setup: Nothing. It's the default — Settings → Browser already selects it.
  • Pros:
    • Zero setup, zero cost, works offline
    • Visible and touchable right in the watch panel
    • Lightest on battery and RAM
    • Takeover taps work directly on the live view
  • Cons:
    • Runs on your phone, so very heavy pages (huge SPAs, WebGL) can be slower than a desktop browser
    • Your phone's IP and user-agent — sites that block mobile/datacenter traffic see a mobile client
  • Best for: Everyday browsing tasks, quick lookups, form filling, reading pages. Start here.

Option B — Cloud Tunnel ​

  • What: A full desktop Chromium running on a server (e.g. a GitHub Actions runner), connected to your phone through your own Cloudflare relay. The app talks to it over CDP (Chrome DevTools Protocol) tunneled over HTTPS/WebSocket.
  • Setup: Deploy the relay worker + runner — full guide: Cloud Tunnel.
  • Pros:
    • Desktop-class rendering and speed for heavy sites
    • Server IP — useful when your mobile IP is blocked or rate-limited
    • The browser keeps running even if the app is backgrounded
  • Cons:
    • Real setup work (Cloudflare Worker + runner + tokens)
    • Needs internet; adds latency per action
    • You operate infrastructure — keep the client token secret
  • Best for: Heavy automation, sites that block mobile clients, long unattended runs.

Recommendation: Use System WebView unless you have a concrete reason not to. It's the default for a reason — instant, free, and good enough for ~90% of tasks. Add the Cloud Tunnel when you hit a site that needs desktop rendering or when you want automation to keep running independently of your phone.


Part 2: Decision engines (the brain) ​

Each browser step is: look at the page → decide one action → do it → repeat. The decision engine is what makes the "decide" step. AgentX tries them in order and falls back gracefully.

  • What: jev-ultrafast-style automation. Instead of sending the whole page to your big main model, AgentX sends a compact element list to a small, fast Choice model that returns exactly one action (~1–2 seconds per step).
  • Setup: Settings → Jev — paste your TypeSafe API key, base URL, and model name. Any Jev-compatible endpoint works.
  • Pros:
    • ~1–2s per action — dramatically faster than model round-trips
    • Cheap — tiny decision model, not your frontier model
    • Fail-open: if Jev isn't configured or errors, the task falls back to the main model instead of breaking
  • Cons:
    • Needs a Jev-compatible key/endpoint (you bring it)
    • Compact element table (~limited chars per step) — can miss context on extremely dense pages
  • Best for: Everyday automation speed. The default choice once configured.

Option B — Drex (Nace AI) ​

  • What: Drex by Nace AI — a wire-compatible alternative to Jev (same API shape, drop-in replacement) that allows roughly ~4x the context per decision step.
  • Setup: Settings → Jev — select Drex as the provider, or paste its base URL. Invite: https://drex.nace.ai/invite/kc6cz8pu
  • Pros:
    • Same fast one-action-per-step loop as Jev
    • ~4x context — sees much larger element tables, better on complex/dense pages
  • Cons:
    • Invite-based access; pricing/tiers set by Nace AI
  • Best for: Complex pages where Jev's compact view misses things — dashboards, dense tables, intricate forms.

Option C — Main model fallback ​

  • What: If neither Jev nor Drex is configured (or the decision call fails), the chat's main model makes the browser decisions.
  • Setup: None — always available.
  • Pros: Works out of the box with zero configuration.
  • Cons: Much slower per step and burns your main model's tokens/context on every click.
  • Best for: Getting started before you've set up Jev/Drex, or as an emergency fallback.

Recommendation: Configure Jev first (fastest, cheapest per action). Switch to Drex when pages get complex and Jev starts missing elements. The main model fallback stays as your safety net — you'll rarely notice it once Jev is set up.


Putting it together ​

Your situationBackendDecision engine
Just installed, want it to workSystem WebViewMain model (then add Jev)
Daily driver, fastest automationSystem WebViewJev
Complex dashboards / dense pagesSystem WebView or TunnelDrex
Site blocks mobile / heavy WebGLCloud TunnelJev or Drex
Long unattended runsCloud TunnelJev

AgentX documentation — community maintained.