Radar / Ideas / ClickSure: portal automations that…

ClickSure: portal automations that can't hallucinate clicks

daily ideamoderateJEV confidence 0.482026-09-25
OutcomeA no-API supplier or partner portal workflow automated in a day, running unsupervised with zero unapproved actions.

The problem

Millions of SMB workflows live behind portals with no APIs: supplier order portals, visa and immigration back-offices, logistics booking sites. Classic RPA cuts out the mid-market: UiPath's minimum bundles and enterprise pricing lead PeerSpot reviewers to say the model 'cuts out the middle market'. Generative computer-use agents are the alternative, but they hallucinate clicks, which is why OpenAI's Operator stayed a limited research preview, was shut down in Aug 2025, and its successor was removed in Aug 2026.

The idea

A UI automation runner where every action is a typed decision drawn from the real UI tree, never generated text. The agent proposes candidate actions, an open action-scoring model ranks them against the current screen state, a Jev decision layer approves or blocks each action before execution, and the whole trace is re-scored after every run. A human approves the action plan once; the runner then executes with a signed, auditable log. Plan-first, gate-every-step, so the failure mode shifts from 'hallucinated click' to 'safe refusal'.

Why now

jev-phone's MIT demo proved the core pattern: Jev picks phone actions directly from the UI tree and never writes a word, so actions cannot be hallucinated. CLM-8B is the first open model that scores candidate agent actions as probabilities up to 9x faster than generation. Qwen Intelligence turned cross-app phone task completion into a full-stack platform. And Operator's shutdown left a real gap for supervised, affordable UI automation.

What it combines

jev-phone-drive-phones-with-jev-a-model-that-never-writes-a-word + qwen-intelligence + clm-8b-open-system-one-model-for-agent-action-scoring + jevals. Phone-first UI-tree action selection plus cross-app orchestration plus open action scoring plus per-trace guardrails. Generation-based agents (browser-use, Operator) guess actions; this stack scores and gates them, which is the property SMBs need to let automations run on someone else's portal unsupervised.

MVP

Weekend MVP: record one desktop browser workflow with browser-use, extract the UI-tree action plan, gate each step through a Jev approve/block decision plus an action scorer, and export the signed run log. Deliberately skip: mobile, self-healing selectors, and the full orchestration UI.

Distribution

B2B2C via agencies and vertical platforms: travel and visa agencies automating supplier portals, 3PL and logistics booking flows, immigration consultancies. One agency pays $299-999/mo and runs it across dozens of client portals; the agency is the wedge into hundreds of SMB workflows.

Why it wins

browser-use is the dominant open framework (80K+ stars) but is generative and DIY; Lindy's Computer Use is bundled into a $49.99/mo general platform; UiPath is enterprise-priced with a heavyweight studio. ClickSure's wedge is decision-gated, plan-first automation with a per-action audit trail, sold to verticals that already own the workflows.

Risks

Biggest risk is portal bot-detection or terms-of-service blocking the runner. The MVP de-risks this by targeting one consented partner portal (an agency's own supplier login) and measuring full-run success rate before broadening to new sites.

Build it with

Repo to start from

clicksure: browser-use workflow recorder plus a typed decision gate (approve/block per action with scoring) and signed run logs for no-API portal workflows.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.