Overview
ClearWeb is an AI shopping guardian. Open it on any checkout, payment, or cookie page, and three AI agents inspect the page, the site's official terms, and the prices in parallel. A larger model combines their results, and every claim is checked against exact on-page evidence before you see it.
It catches the tricks checkouts use to push up the price:
- Protection plans and add-ons that are already ticked
- Fees that only appear at the final step
- Free trials that renew automatically
- Cashback presented as if it were a discount
- EMI plans with hidden charges
- Cookie banners that make "accept all" the easy choice
It is built for online shoppers, with first-class support for Indian e-commerce and payment flows: UPI, EMI, and cashback.
Features
- Multi-agent analysis: three specialist agents (page, official terms, prices) run concurrently, and an orchestrator model synthesizes their results.
- Evidence-verified findings: every amount and claim must match text captured from the page, down to word and number boundaries. Unsupported claims never reach the user as facts.
- Charge breakdown: amounts are grouped into what you pay now, fees and taxes, items, discounts, later benefits such as cashback, and list prices (MRP).
- One-click fixes: with your permission, ClearWeb switches off optional add-ons and marketing cookies, verifies the result, and offers Undo.
- Three safety modes: Observe (read-only), Assist (acts only after a click), and Auto-fix (enabled per site, limited to removing verified optional extras).
- Live re-analysis: page changes are hashed and debounced at 750 ms, so the analysis stays current while cosmetic changes cost nothing.
- Workspace UI: a draggable, resizable React panel with Page, Compare (up to four tabs), and Activity views, follow-up chat, a Ctrl/Cmd+K command palette, "Show on page" highlighting, light and dark themes, and reduced-motion support.
How it works
- Capture. The content script builds a redacted outline of every visible row and control on the page.
- Triage. A fast Nemotron Lightning model selects the relevant blocks by ID only, so it can never introduce text. The selection is packed into a compact 6 KB snapshot, with totals, fees, and preselected options always kept.
- Analyze. Three Nemotron Lightning workers analyze the page, the official terms, and the prices in parallel. Progress streams live to the UI as typed NDJSON events.
- Synthesize. Nemotron Ultra reconciles the workers' findings, answers follow-up questions, and proposes actions from a fixed allowlist.
- Verify. A deterministic verifier checks every amount with Decimal arithmetic, matches every quote to captured page text, and authorizes each action before the extension runs it.
Tech stack
- Extension: TypeScript, React 19, Chrome Manifest V3 (service worker, content scripts, side panel, chrome.storage), esbuild
- Backend: Python, FastAPI, Pydantic v2, httpx (async), Uvicorn
- AI: NVIDIA Nemotron 3.5 Lightning (workers and triage) and Nemotron 3 Ultra (orchestrator) on Nebius, plus optional domain-restricted Tavily search
- Data: SQLite for the cost ledger and result cache
- Infrastructure: self-hosted through Cloudflare Tunnel, with TypeScript API contracts generated from OpenAPI
- Testing: pytest and Playwright
Engineering highlights
The AI never gets the last word. The models propose findings; deterministic code decides what is shown. Quotes must match captured text exactly, a negative amount can never count as proof of a charge, and both 1,234.56 and 1.234,56 number formats are parsed correctly.
Model output can't touch the page. The extension never executes code or selectors that come from a model. Actions come from a fixed allowlist, can only switch things off, and are checked after they run. ClearWeb never pays, places orders, or accepts contracts.
Privacy by design. Emails, UPI IDs, card and payment fields, addresses, long digit sequences, tokens, form values, and cookies are redacted before anything leaves the browser. Raw HTML is never sent, and URLs are stripped of query strings.
Built to run cheaply and reliably. A SQLite ledger reserves the cost of every model call before dispatch, under a hard spending cap; a full page analysis costs about three cents. Per-page locks, per-tab session state, request deduplication, retries with backoff, and a rule-based fallback keep analysis stable across many tabs.
Production-hardened API. Every request needs an API key, with per-client rate limiting, 64 KB request limits, an LRU result cache, and timeouts on every model call.