TPEmist

Point One Percent — pop-pay

Community TPEmist
Updated

Project Point One Percent - Agent Pay [pop-pay] is a payment guardrail and one-time flow protocol specifically designed for Agentic AI (e.g., Claude Code, OpenClaw). It enables agents to handle financial transactions safely without risking unlimited exposure of human-controlled credit cards.

Point One Percent — pop-pay

it only takes 0.1% of Hallucination to drain 100% of your wallet.

The runtime security layer for AI agent commerce.

Your card never enters the agent's context. One hallucinated prompt can't drain a wallet it can't see.

pop-pay is an open-source (MIT) runtime security layer that protects AI agents during online purchases. It works with OpenClaw, NemoClaw, Claude Code, OpenHands, and any MCP-compatible framework.

Architecture: Five Security Primitives

Primitive What it does
Context Isolation Layer Card credentials are injected directly into the browser DOM via CDP — they never enter the agent's process or LLM context window. Prompt injection can't steal what the agent doesn't have.
Intent Verification Engine Hybrid keyword + LLM guardrail evaluates whether a purchase should happen — not just whether it can. 95% accuracy on 20-scenario benchmark.
Human Trust Anchor Configurable human-in-the-loop approval for high-value or unrecognized transactions.
Zero-Knowledge Card Surface Agent only sees masked tokens (****-4242). Real data is stored in an AES-256-GCM encrypted vault.
Ephemeral Authorization Scope Each payment approval is single-use with TOCTOU domain guard — an approved session can't be redirected to a malicious merchant.

See THREAT_MODEL.md for the full STRIDE analysis and COMPLIANCE_FAQ.md for enterprise compliance details.

Guardrail Benchmark

Layer Score Notes
Keyword only 14/20 (70%) Fast, zero-cost, catches obvious violations
Hybrid (Keyword + LLM) 19/20 (95%) LLM resolves 5 of 6 keyword failures
Feature AgentPayy AgentWallet Prava pop-pay
Enforcement Mock alert() Rule-based Spending limits Semantic validation
Intent check None Agent-provided text None Context-aware LLM
Injection-proof No No No Yes

Two Deployment Modes

BYOC — Bring Your Own Card (Local)

The agent never receives the true card number — it only sees ****-4242. When checkout is reached, the Context Isolation Layer attaches via CDP, traverses all cross-origin iframes (Stripe Elements, Adyen, etc.), and injects credentials directly into the DOM. Runs entirely on your machine — no SaaS, no login, no external account.

Enterprise — Stripe Issuing

For cloud-hosted AI fleets: programmatically issue single-use virtual cards via Stripe API, with per-agent budgets and full audit trails.

Ecosystem Position

pop-pay is the agent's Policy Enforcement Point — it evaluates, approves, and injects. It does NOT navigate websites or solve CAPTCHAs — that's the browser agent's job.

The Handshake: How Point One Percent and Browser Agents Work Together

The real power emerges when Point One Percent is paired with a browser automation agent (e.g., OpenHands, browser-use, Skyvern). The workflow is a clean division of labor:

1. [Browser Agent]  Navigates to a site, scrapes product info, reaches checkout.
        │
        │  (Hit a paywall / payment form)
        ▼
2. [Browser Agent → POP MCP]  Calls request_virtual_card(amount, vendor, reasoning)
        │
        │  (Point One Percent evaluates: budget OK? vendor approved? no hallucination?)
        ▼
3. [POP]  Issues a one-time virtual card (Stripe mode) or mock card (dev mode).
            Full card credentials handled only by the local trusted process —
            never exposed to the agent or LLM context.
        │
        ▼
4. [POP]  Injects real credentials into the checkout form via CDP.
            The agent receives only a transaction confirmation — no card details.
        │
        ▼
5. [Browser Agent]  Clicks the submit button to complete the transaction.
        │
        ▼
6. [The Vault]  Dashboard logs the transaction. Card is immediately burned.

Supported Integrations

Integration path Works with
MCP Tool Claude Code, OpenClaw, NemoClaw, OpenHands, any MCP-compatible host
Python SDK Custom Playwright, browser-use, Skyvern, Selenium, gemini-cli

Any browser-capable agent (Claude Code, OpenClaw, browser-use, Skyvern, etc.) gets full CDP injection — card is auto-filled into the payment form, the agent only ever sees the masked confirmation (****-****-****-4242). See the Integration Guide for setup instructions and System Prompt templates.

4. Installation

Shell note: [...] is special syntax in zsh and bash — always wrap the package name in quotes.

# Core only (keyword guardrail + mock provider, zero external dependencies)
pip install "pop-pay"

# Claude Code / MCP integration
pip install "pop-pay[mcp]"

# Claude Code + CDP injection (BYOC)
pip install "pop-pay[mcp,browser]"

# With LLM-based guardrails (supports OpenAI, Ollama, vLLM, OpenRouter)
pip install "pop-pay[mcp,llm]"

# With Stripe virtual card issuing
pip install "pop-pay[stripe]"

# With LangChain integration
pip install "pop-pay[langchain]"

# Full installation (all features)
pip install "pop-pay[all]"

5. Quick Start for OpenClaw / NemoClaw / Claude Code / OpenHands

If you're using OpenClaw, NemoClaw, Claude Code, OpenHands, or any MCP-compatible agentic framework, you can get Point One Percent running in under 2 minutes:

Step 1: Set Up Environment & Install

# Create a dedicated directory and virtualenv
mkdir ~/pop-pay && cd ~/pop-pay
python3 -m venv .venv && source .venv/bin/activate

# Install — quotes required for zsh/bash
pip install "pop-pay[all]"

Contributing / local development? See CONTRIBUTING.md for the git clone + uv sync path.

Step 1b: Initialize the Credential Vault

v0.6.0+: Credentials are stored in an AES-256-GCM encrypted vault — no plaintext .env required.

pop-init-vault

This will prompt for your card credentials (input is hidden), encrypt them into ~/.config/pop-pay/vault.enc, and securely wipe any existing .env. The MCP server auto-decrypts the vault at startup — no extra steps needed per session.

Passphrase mode (stronger — protects against agents with shell access):

pop-init-vault --passphrase   # one-time setup
pop-unlock                     # run once before each MCP server session

pop-unlock derives the key from your passphrase and stores it in the OS keyring. The MCP server reads it automatically at startup.

Security levels (lowest → highest):

Mode Protects against
.env file (legacy) Nothing — plaintext on disk
Vault, machine key, OSS source File-read agents
Vault, machine key, pip install pop-pay File-read agents + casual shell inspection
Vault + passphrase File-read agents + shell agents
Stripe Issuing (commercial) All local threats — no credentials stored

Policy & non-credential config (allowed vendors, spending limits, CDP URL) is still read from ~/.config/pop-pay/.env. Only card credentials moved to the vault.

Step 2: Launch Chrome & Get MCP Commands

pop-launch --print-mcp

This launches Chrome with CDP enabled and prints the exact claude mcp add commands to run.

Step 3: Add to Claude Code

Choose your platform and follow the dedicated setup guide:

Platform Setup Guide
Claude Code (BYOC + CDP injection, recommended) Integration Guide §1
Python script / gemini-cli Integration Guide §2
Playwright / browser-use / Skyvern Integration Guide §3
OpenClaw / NemoClaw Integration Guide §4 / OpenClaw Skill
OpenHands Add python -m pop_pay.mcp_server to your mcpServers config

Step 4: Configure Policy

Edit ~/.config/pop-pay/.env (see Step 1b). Key variables:

Variable Default Description
POP_ALLOWED_CATEGORIES ["aws","cloudflare"] Vendors the agent is allowed to pay — see Categories Cookbook
POP_MAX_PER_TX 100.0 Max $ per transaction
POP_MAX_DAILY 500.0 Max $ per day
POP_BLOCK_LOOPS true Block hallucination/retry loops
POP_AUTO_INJECT false Enable CDP card injection
POP_GUARDRAIL_ENGINE keyword Guardrail engine: keyword (zero-cost, default) or llm (semantic, two-layer) — see Guardrail Mode
POP_BILLING_FIRST_NAME / POP_BILLING_LAST_NAME (empty) Auto-fill name fields on checkout pages
POP_BILLING_EMAIL (empty) Auto-fill email
POP_BILLING_PHONE (empty) E.164 format — auto-fill combined phone input
POP_BILLING_PHONE_COUNTRY_CODE (empty) ISO code ("US") or dial prefix ("+1") — fills country code dropdown; national number auto-derived
POP_BILLING_STREET / POP_BILLING_CITY / POP_BILLING_STATE / POP_BILLING_COUNTRY / POP_BILLING_ZIP (empty) Auto-fill address fields; state and country matched fuzzily against dropdown options
POP_ALLOWED_PAYMENT_PROCESSORS [] Extra third-party payment processor domains to trust (pop-pay ships with 20 built-in)
POP_WEBHOOK_URL (empty) Webhook URL for Slack/Teams/PagerDuty notifications

After editing .env, fully close and reopen Claude Code. The MCP server loads configuration at startup — !claude mcp list alone is not sufficient to pick up .env changes.

Guardrail Mode: Keyword vs LLM

Point One Percent ships with two guardrail engines. You switch between them with a single env var:

keyword (default) llm
How it works Blocks requests whose reasoning string contains suspicious keywords (e.g. "retry", "failed again", "ignore previous instructions") Sends the agent's reasoning to an LLM for deep semantic analysis
What it catches Obvious loops, hallucination phrases, prompt injection attempts Subtle off-topic purchases, logical inconsistencies, policy violations that keyword matching misses
Cost Zero — no API calls, instant One LLM call per request_virtual_card invocation
Dependencies None Any OpenAI-compatible endpoint
Best for Development, low-risk workflows, cost-sensitive setups Production, high-value transactions, untrusted agent pipelines

Tip: keyword mode requires no extra config. To enable LLM mode, see the full configuration reference in the Integration Guide §1.

Step 4: Use It

Your agent now has access to two tools:

Tool When to use
request_purchaser_info Billing/contact info page (name, email, phone, address) — no card fields visible yet
request_virtual_card Payment page — card fields are visible. Prompt injection scan runs automatically inside this call.

Single-page checkout (e.g. Wikipedia donate): agent calls request_virtual_card.Two-page checkout (e.g. billing info → payment): agent calls request_purchaser_info first, then request_virtual_card.

When it encounters a paywall:

Agent: "I need to purchase an API key from AWS for $15 to continue."
[Tool Call] request_virtual_card(amount=15.0, vendor="AWS", reasoning="Need API key for deployment")
[POP] Payment approved. Card Issued: ****4242, Expiry: 12/25, Amount: 15.0
Agent: "Purchase successful, continuing workflow."

If the agent hallucinates or tries to overspend:

Agent: "Let me retry buying compute... the previous attempt failed again."
[Tool Call] request_virtual_card(amount=50.0, vendor="AWS", reasoning="failed again, retry loop")
[POP] Payment rejected. Reason: Hallucination or infinite loop detected in reasoning

6. Core Components

The Vault

A local visualization console powered by Streamlit and SQLite (pop_state.db). The Vault allows humans to:

  • Monitor all issued seals and agent spending activity in real-time.
  • Monitor global budget utilization.
  • Audit rejection logs from semantic guardrails.

The Seal

Virtual, single-use payment credentials with built-in enforcement:

  • Daily Budget Limit Enforcement: Automatically blocks any request that would exceed the predefined daily spending cap.
  • Burn-after-use Interception: Ensures that once a virtual card is used, it is immediately invalidated, preventing replay attacks or unauthorized recurring charges.

Semantic Guardrails

Point One Percent provides two modes of intent evaluation. Both are controlled by POP_GUARDRAIL_ENGINE in your .env (see §5 Step 3 for full configuration).

  1. Keyword mode (POP_GUARDRAIL_ENGINE=keyword, default): The GuardrailEngine scans the agent's reasoning string for suspicious phrases associated with loops or hallucinations (e.g., "retry", "failed again", "ignore previous"). Zero dependencies, zero latency, zero cost. Recommended as the starting point for all setups.

  2. LLM mode (POP_GUARDRAIL_ENGINE=llm): The LLMGuardrailEngine sends the agent's reasoning to an LLM for deep semantic analysis, catching subtler misuse that keyword matching would miss — such as off-topic purchases or logically inconsistent justifications. Supports any OpenAI-compatible endpoint: OpenAI, Ollama (local), vLLM, OpenRouter, and more.

7. Security Statement

Security is a first-class citizen in Point One Percent. The SDK masks card numbers by default (e.g., ****-****-****-4242) when returning authorization results to the agent. This prevents sensitive payment information from leaking into agent chat logs, model context windows, or persistent logs, ensuring that only the execution environment handles the raw credentials.

v0.6.0 defense-in-depth hardening:

Layer Defense
Encrypted vault Card credentials stored as AES-256-GCM ciphertext (vault.enc); plaintext never touches disk after pop-init-vault
Passphrase mode Key derived from user passphrase via PBKDF2 (600k iterations); stored in OS keyring — agents with shell access cannot derive the key
Database SQLite only stores masked card (****-4242); card_number and cvv columns removed entirely
Injection-time TOCTOU guard Domain verified against guardrail-approved vendor at the moment of injection — prevents redirect-to-attacker attacks
Repr redaction VirtualSeal.__repr__ always emits ****-REDACTED; credentials cannot leak via logs or tracebacks
Core dump prevention MCP server disables core dumps at startup (RLIMIT_CORE=0)
Process isolation Agent communicates via MCP JSON-RPC as a separate process — cannot access MCP server memory or env vars through the protocol

See SECURITY.md for the full threat model, red team results, and documented limitations.

8. The Vault Dashboard

The Vault is your real-time monitoring console for all agent payment activity.

Starting the Dashboard

cd Point-One-Percent
uv run streamlit run dashboard/app.py
# Dashboard opens at http://localhost:8501

Dashboard Layout

Section Description
Sidebar: Max Daily Budget slider Adjust the displayed budget cap for visualization (does not affect backend policy — backend policy is configured via env vars or SDK)
Today's Spending Total amount spent by agents today
Remaining Budget How much budget is left for the day
Budget Utilization Visual progress bar showing spend %
Issued Seals & Activity Full table of all payment attempts (approved + rejected) with seal ID, amount, vendor, status, and timestamp
Rejected Summary Filtered view showing only rejected/blocked attempts for quick auditing

Tips

  • Click Refresh Data in the sidebar to pull latest activity from the database.
  • The dashboard reads from pop_state.db — the same database the SDK writes to. Keep both running simultaneously for live monitoring.
  • Each row in the table corresponds to a single request_virtual_card call from an agent.

9. Python SDK Quickstart

Integrate Point One Percent into your custom Python or LangChain workflows in just a few lines:

from pop_pay.client import PopClient
from pop_pay.providers.stripe_mock import MockStripeProvider
from pop_pay.core.models import GuardrailPolicy

# Define your safety policy
policy = GuardrailPolicy(
    allowed_categories=["API", "Cloud", "SaaS"],
    max_amount_per_tx=50.0,
    max_daily_budget=200.0,
    block_hallucination_loops=True
)

# Initialize the client with keyword-only guardrails (default)
client = PopClient(
    provider=MockStripeProvider(),
    policy=policy,
    db_path="pop_state.db"
)

# Use with LangChain Tool
from pop_pay.tools.langchain import PopPaymentTool
tool = PopPaymentTool(client=client, agent_id="agent-01")

For LLM guardrail engine setup and the full provider reference, see Integration Guide §2.

10. Payment Providers: Stripe vs Mock

Without Stripe (Default — Mock Provider)

By default, Point One Percent uses the MockStripeProvider which simulates virtual card issuance. This is perfect for:

  • Development and testing — no real money involved
  • Demo and evaluation — see the full flow without any API keys
  • Hackathons — get a working prototype in minutes

Mock cards are fully functional within the system (budget tracking, burn-after-use, guardrails all work), but they are not real payment instruments.

BYOC — Bring Your Own Card (Hacker Edition)

For developers who want to use their own physical credit card with Point One Percent without a Stripe account. The LocalVaultProvider reads card credentials from environment variables and injects them into browser payment forms via CDP — the raw PAN is never exposed to the agent.

Add to your ~/.config/pop-pay/.env:

POP_BYOC_NUMBER=4111111111111111   # Your real card number
POP_BYOC_CVV=123
POP_BYOC_EXP_MONTH=12              # Expiry month, e.g. 04
POP_BYOC_EXP_YEAR=27               # Expiry year, e.g. 31
POP_AUTO_INJECT=true

Then restart Claude Code. The MCP server will automatically use LocalVaultProvider.

Provider priority (high → low): Stripe Issuing → BYOC Local → Mock.

If POP_STRIPE_KEY is set, Stripe takes precedence. If POP_BYOC_NUMBER is set (but no Stripe key), LocalVaultProvider is used. If neither is set, MockStripeProvider is used for development.

Security note: Never commit real card numbers to version control. Always use .env (which is .gitignored) or a secrets manager. The CDP injection ensures the full card number is only handled by the local trusted process, never by the LLM.

For Python SDK usage of each provider, see Integration Guide §2.

With Real Stripe Issuing

To issue real virtual credit cards through Stripe Issuing:

Prerequisites:

  1. A Stripe account with Issuing enabled (requires application approval)
  2. Your Stripe secret key (sk_live_... or sk_test_...)

Option A: Via Environment Variable (for MCP Server)

export POP_STRIPE_KEY=sk_live_your_stripe_key_here
python -m pop_pay.mcp_server
# The MCP server will automatically use StripeIssuingProvider

What Stripe Issuing does:

  • Creates a real Stripe Cardholder (POP Agent)
  • Issues a virtual card with a spending limit matching the approved amount
  • Returns masked card details (last 4 digits only) to the agent
  • All Stripe errors are caught and returned as rejection reasons

CDP injection limitation with Stripe Issuing: The Stripe Issuing API returns only the last 4 digits of the card number for security reasons. CDP auto-injection (POP_AUTO_INJECT=true) requires the full PAN and therefore does not work with Stripe Issuing. Use BYOC (POP_BYOC_NUMBER) if you need CDP injection; use Stripe Issuing if you need a real card and will handle form submission yourself.

Note: Stripe Issuing is a premium Stripe product that requires approval. For most development and demo use cases, the Mock provider is sufficient.

MCP Server · Populars

MCP Server · New

    Prismer-AI

    Signet

    Cryptographic action receipts for AI agents — sign, audit, verify.

    Community Prismer-AI
    ctxray

    ctxray

    See how you really use AI — X-ray your AI coding sessions locally

    Community ctxray
    vkhanhqui

    figma-mcp-go

    Figma MCP for free users — no rate limits, full read/write, text to designs, designs to code

    Community vkhanhqui
    cerul-ai

    Cerul

    The video search layer for AI agents. Search video by meaning — across speech, visuals, and on-screen text.

    Community cerul-ai
    GuyMannDude

    ⚡ Mnemo Cortex v2.2

    Open-source memory coprocessor for AI agents. Persistent recall, semantic search, crash-safe capture. No hooks required.

    Community GuyMannDude