Wallet Guard Agent

AI agents are starting to hold wallets, sign transactions, and move funds on their own — but they're doing it without any shared sense of risk. A single honeypot token or an unlimited approve call is enough to drain a wallet, and right now there's no common layer an agent can check before it acts. Humans have Etherscan tabs and gut instinct; agents have nothing.

SecurityGuard is our answer: a security layer built for agents, not people. It's an AI-powered risk oracle that any wallet, trading bot, or DeFi agent can call before signing — sending back a simple, machine-readable verdict: Approve, Warn, or Block, backed by a calibrated risk score and a clear explanation.

Under the hood it's a LangGraph agent that pulls from multiple blockchain and security data sources in parallel, then uses Claude to reason over the combined signal the way a human security researcher would — but in under three seconds, and at a fraction of the cost. It's multi-chain, pay-per-use ready via OKX Web3, and built to be listed and called directly from OKX.AI.

Our bet is simple: as more of the economy moves to agents acting autonomously on-chain, "did anyone check this first?" becomes a question every agent needs answered — automatically, every time, before the money moves. SecurityGuard is that check.

Designed and built SecurityGuard Agent end-to-end as a LangGraph workflow: input validation → parallel data fetching → parallel security analysis → LLM reasoning → structured output formatting.

• Implemented the core FastAPI service (app.py) with a stable /v1/analyze contract, /health, and /docs, plus API-key auth and rate limiting.

• Built the LangGraph agent (agent.py) with a heuristic fallback path so the product stays demoable even without LLM access.

• Integrated multiple blockchain and security data sources: Web3.py/RPC for on-chain data, honeypot.is and GoPlus for token/address risk signals, explorer APIs for contract verification and tx history, plus custom bytecode heuristics (SELFDESTRUCT/DELEGATECALL detection) and unlimited-approve calldata detection.

• Added OKX Web3 enrichment and a pay-per-use payment verification flow (okx_tools.py), designed to be Agent Trade Kit–compatible.

• Wired up Claude Sonnet 5 as the reasoning layer to turn raw findings into a calibrated 0–100 risk score, Low/Medium/High/Critical rating, and Approve/Warn/Block recommendation with a human-readable explanation.

• Built a Gradio demo UI (demo.py) with preset scenarios (safe USDT vs. unlimited-approve phishing) for fast, repeatable demoing.

• Set up SQLite-backed audit logging, containerized the service (Dockerfile), and prepared Railway/Render deploy configs for one-click hosting.

• Recorded a demo comparing a clean token check against a phishing-style transaction to show the scoring and Block logic working end-to-end.

Not yet raised. We're self-funded through the hackathon and plan to pursue pre-seed funding afterward to expand chain coverage, security integrations, and move SecurityGuard toward production-grade infrastructure.