Agent Shopping Safe Checkout
pawel-cell/agent-shopping-safe-checkout
Safe-checkout pattern for agents that buy online: isolated Chrome, virtual card and a human approval gate
Agent Shopping Safe Checkout is a Hermes skill and AGENTS.md pattern that lets an AI agent complete real web purchases with three independent guardrails: browser isolation, payment isolation and a human gate.
What Agent Shopping Safe Checkout does
The first guardrail is a fresh Chrome --user-data-dir, never your daily profile. The second is a one-time, merchant-locked or low-cap virtual card. The third is a human gate: the agent must receive an explicit submit order phrase from you before the final click. The README is clear about which layer matters most. The hard security boundary is the virtual card's spend cap, merchant lock and expiration, while prompt rules are a soft control, because any merchant page or popup chat may try to re-instruct the agent.
Two files carry the pattern. SKILL.md is a Hermes Agent skill meant for ~/.hermes/skills/productivity/agent-shopping-safe-checkout/, and it loads automatically on shopping triggers. AGENTS.md is a drop-in instruction file for Claude Code, OpenAI Codex CLI, Cursor, Cline, Aider or any model with shell access, and Cursor users can paste it into .cursorrules. The pattern combines Browser Harness, a thin CDP bridge to Chrome, with Stripe Link CLI one-time cards for US Link accounts, or Privacy.com, Revolut and Wise cards elsewhere.
Key features
- Browser isolation with a fresh Chrome user data directory
- Payment isolation with one-time, merchant-locked or low-cap virtual cards
- Human gate requiring an explicit submit order phrase before checkout
- Hermes SKILL.md that auto-loads on shopping triggers
- AGENTS.md for Claude Code, Codex CLI, Cursor, Cline and Aider
When to use it
- Letting an agent fill a real shopping cart while you approve the final purchase
- Limiting the damage if a merchant page tries to redirect the agent
- Reusing one checkout policy across several coding agents
Who it is for: People who want an AI agent to buy things online without exposing their main browser profile or bank card.
How it fits with Hermes Agent
It ships a Hermes Agent SKILL.md for the ~/.hermes/skills tree and also an AGENTS.md for other agents, so Hermes is one of several supported agents.
Requirements: A one-time, merchant-locked or low-cap virtual card, and Browser Harness driving a fresh Chrome profile
FAQ
What is Agent Shopping Safe Checkout?
Agent Shopping Safe Checkout is a security pattern for AI agents that make real web purchases. It combines an isolated browser, a virtual card with a spend cap, and a human confirmation phrase.
Does Agent Shopping Safe Checkout work with Hermes Agent?
Yes, it includes a SKILL.md for Hermes Agent that you place under ~/.hermes/skills/productivity/. A separate AGENTS.md covers Claude Code, Codex CLI, Cursor, Cline and Aider.
Is Agent Shopping Safe Checkout free and open source?
The README states MIT, but GitHub detects no license file in the repository, so confirm with the author before reuse.
Similar security for Hermes Agent
All securityDaily read-only supply-chain scans for a Hermes Agent host using Perplexity's Bumblebee
nativ3ai Hermes PayGuardSafe-by-design USDC and x402 payment plugin for Hermes Agent with human approval for larger transfers
jayelbotvibe-web Hermes Pentest LabKali-native pentesting lab with an MCP tool server, guardrails enforced in code and PDF reports
Tranquil-Flow Hermes-AegisSecurity layer for Hermes Agent: a MITM proxy that blocks secret leaks and dangerous commands
MahdiHedhli HermesUltraCodeHermes plugin that has a second-lab model review each task before a subagent starts writing code
intentframe IntentFrame for Hermes AgentIntentFrame security plugin that checks Hermes terminal, code, file and cron tool calls against policy
Related guides: How to run Hermes Agent securely