Canary

Canary homepage with search bar
Guardian running trust checks in real-time
Scam site blocked, trust score 17
Agent bypass attempt caught
Amazon approved, green checks
PayPal order created successfully
AG Grid analytics dashboard
Architecture diagram (two agents, five checks)

Inspiration

AI shopping agents optimize for price — that's their job. But without a trust layer, they'll happily buy from a 3-day-old domain selling Sony headphones at 84% off market price. The payment goes through. The money is gone. No human ever verified if the merchant was real.

We asked: what if every agentic payment had a canary?

What it does

Canary is a multi-agent trust verification system that sits between an AI shopping agent and PayPal checkout. When an AI agent wants to buy something:

Shopper Agent finds the cheapest option (intentionally naive — optimizes for price only)
Guardian Agent challenges every claim with 5 real-time trust signals:
- Domain age via WHOIS (new domains = risky)
- SSL certificate validation (self-signed = suspicious)
- Price analysis via Channel3 API (too cheap = bait)
- Review authenticity via Claude AI (fake review pattern detection)
- Business registry check (TLD heuristics + domain patterns)
A weighted Trust Score (0–100) determines: APPROVED, BLOCKED, or ASK USER
A Bypass Monitor catches the agent if it tries to skip Guardian and call PayPal directly

The adversarial two-agent design means the system doesn't just verify — it challenges. Guardian treats Shopper's claims as assumptions to be proven, not facts to be accepted.

How we built it

Backend: Python + FastAPI with SSE streaming for real-time event delivery
Agents: Two adversarial agents (Shopper + Guardian) coordinated by an Orchestrator
Payments: PayPal Agent Toolkit (Orders API v2, sandbox mode)
Site Browsing: Kernel SDK cloud browsers — agents browse real websites, not cached data
Price Data: Channel3 API (100M+ products) for market price comparison
AI Intelligence: Anthropic Claude for review analysis and assumption verification
Frontend: Next.js 14 + TypeScript + Tailwind CSS + Framer Motion
Analytics: AG Grid-powered dashboard for transaction history and trust analytics
Database: SQLite for transactions, merchants, and bypass attempts

Challenges we ran into

PayPal Agent Toolkit had an undocumented import path change that took debugging to find
Agent bypass detection — designing a state machine that catches an AI agent trying to circumvent user consent without false positives required careful state tracking
Real-time streaming — bridging synchronous agent execution with async SSE required an asyncio.Queue event bridge pattern
Stale React closures — useCallback captured old state, breaking the canary toggle in auto-start mode

Accomplishments we're proud of

The adversarial agent design actually works — Guardian catches scams that Shopper happily approves
The bypass monitor catches unauthorized PayPal API calls with zero false positives
Real trust signals from real APIs — no mocked data, no hardcoded scores
A real PayPal sandbox order is created for approved purchases

What we learned

Trust in agentic commerce isn't just about verifying merchants — it's about verifying your own agent's assumptions
The adversarial multi-agent pattern produces better outcomes than a single agent trying to be both smart and safe
Price anomaly detection needs market data to be meaningful — "cheap" is relative