{"id":47234695,"title":"Agent Pro – Automate your desktop from your phone (no setup)","points":1,"user":"ypadamat","time":1772554785,"time_ago":"5 hours ago","type":"link","content":"

We built an AI agent that lets you automate your desktop from a phone app. \nNo servers to manage, no npm install, no configuration. Sign in with Cleer, \ndescribe what you want automated, and it happens.

Why we built this:\nOpenClaw is great, but it has a 2-hour setup tax (Node.js, gateway config, \nskill vetting). Most people never make it past npm install. We wanted zero \nsetup.

The technical problem we solved:\nDesktop automation usually relies on screenshot + LLM vision. The problem: \nLLM vision gives wrong coordinates, especially on high-DPI displays. On a \nRetina Mac, "click at (523, 412)" might be off-screen because of 2x scaling.

Our approach:\n- DOM injection (live webpage elements with native coordinates)\n- A11y tree snapshots (native app UI, pixel-perfect)\n- Screenshot as fallback (only when needed)\n- JavaScript reads window.devicePixelRatio, eliminates coordinate guessing

Result: ±2px accuracy instead of ±50px. It actually works.

The workflow:\nPhone: "automate my expense report filing"\nAgent Pro: Figures out what you need → Runs it on your desktop → Sends \nresult back to phone\nYou: Get confirmation in ~30 seconds

No servers running on your laptop. No daemon crashes. No "did my Tailscale \ntunnel drop?" Cloud-managed, syncs across all your devices. (Agent runs on your PC but memory is stored with your account, we can never see it)

Technical highlights:\n- Hierarchical perception (DOM > A11y > Screenshot > Blind mode)\n- Lane queue system (serial by default, prevents race conditions)\n- Reflection engine (detects loops, forces strategy pivots)\n- API failover (Grok + MiniMax with health tracking)\n- MiniMax vision pipeline (text-only M2.5 → vision analysis via minimax-01 \n→ structured JSON)\n- DPI/multi-monitor support (native coordinate space, no scaling guessing)

What we didn't do:\n- We don't have 5,700 skills like OpenClaw (that's OpenClaw's strength)\n- We didn't build a 23-channel gateway (not our use case)

What we did do:\n- Made it work out of the box (sign in, done)\n- Made it accurate (DOM extraction > screenshot vision)\n- Made it mobile compatible (phone control is the whole point, can still use via desktop tho)\n- Made it non-flaky

We're launching soon for Cleer users. Happy to answer questions.

https://solnetex.com","url":"item?id=47234695","comments":[],"comments_count":0}