Documentation v1.24
Last updated 2026-05-25 Give feedback

MagicBrowse

MagicBrowse helps agents use legacy commerce websites by combining browser automation, checkout-specific behavior, and a growing site memory.

  • Browser-use agent purpose-built for real e-commerce navigation and checkout flows.
  • LLM planner with vision plus a high-speed DOM navigator for checkout execution.
  • Stops at login, payment, identity, CAPTCHA, and other Memory boundaries.
  • Learns from successful runs through site memory, runtime hints, and checkout traces.
MagicBrowse checkout trace

Commerce-First Browser Use

General browser agents are broad and brittle around checkout. MagicBrowse narrows the problem: search result pages, product pages, carts, guest checkout, provider handoff, form observation, and sensitive-step detection. The active package is `@mercuryo-ai/magicbrowse`, and the active CLI surface is `@mercuryo-ai/magicbrowse-cli`.

Planner And Navigator

MagicBrowse separates planning from low-level browser movement. The planner reasons over the task, visual state, and commerce-specific policies. The navigator performs DOM-oriented actions and reports structured observations. This lets MagicPay keep the agent fast on ordinary browsing while enforcing hard stops before sensitive entry.

TypeScript
import { act, launch } from '@mercuryo-ai/magicbrowse'; const session = await launch({  headless: false,  url: 'https://www.booking.com',}); const result = await act({  sessionId: session.id,  goal: 'Find a Hilton stay from May 10 to May 12 and stop before sensitive guest or payment data.',}); if (result.status === 'needs_handoff') {  await createMagicPayRequest(result.handoff);}

Sensitive-step Handoff

MagicBrowse should not type passwords, card values, passport details, OTPs, or final payment confirmations directly from model context. When it observes a Memory-managed form, MagicPay creates a request, resolves the approved artifact, fills through the trusted Memory-managed boundary, and returns only the continuation state the browser agent needs.

Site Memory

MagicBrowse teaches itself after each successful checkout. Successful commerce traces can be converted into site-specific hints: useful URLs, checkout milestones, button labels, form shapes, provider handoff patterns, and recovery notes. These hints speed up future runs and reduce repeated exploration, but they never override live page state, policy, or user approval. Over time, this becomes an agentic knowledge base of internet commerce: a practical memory of how real merchants, booking sites, marketplaces, and checkout flows work for AI agents.

MagicPay public documentation.