DeFi Guard MCP

The safety layer that checks a DeFi transaction or token before your agent (or you) signs — on Base L2. Sits in front of execution MCPs (it pairs with them, it doesn't compete): honeypot detection, owner-power scans, approval-drain checks, position health, and executable prices — all from live on-chain state.

No API keys required — works out of the box against public Base RPCs (bring your own RPC for speed via BASE_RPC_URL). Read-only: it never holds keys, signs, or submits.

Guard-before-signing tools

Tool	What it answers
`token_safety_screen`	"Is this token safe to buy/approve before I sign?" — honeypot detection (can you actually sell it back?), real round-trip cost (fees + tax both ways), and whether ownership is renounced (a live owner can often change taxes / pause / mint). One risk verdict.
`scan_dangerous_capabilities`	"What can the owner do to me?" — scans the deployed bytecode for owner-only powers: `mint`, `pause`, `blacklist`, adjustable fees/taxes, max-tx limits, trading toggles, proxy `upgradeTo`. Flags the capability, no explorer key needed.
`approval_risk`	"Is this approval dangerous?" — reads the live allowance an owner granted a spender, flags unlimited approvals (the allowance-drain vector) and whether the spender is a contract or an EOA. Current exposure = what could be pulled right now.

Data tools

Tool	What it answers
`aave_position_health`	"Is this Aave V3 position safe?" — live health factor, collateral/debt in USD, LTV, liquidation threshold, and a plain risk level (`healthy` / `elevated` / `critical` / `LIQUIDATABLE`).
`quote_swap`	"What would this swap actually return right now?" — exact-input quote via Uniswap V3 QuoterV2, best of all 4 fee tiers, with gas estimate. Executable price, not an oracle.
`token_risk_snapshot`	"Can I get out of this token?" — ERC-20 metadata + real market depth measured by round-trip quotes (WETH → token → WETH) at two sizes. High round-trip loss = thin or trapped liquidity, whatever the chart says.

Why round-trip depth instead of "liquidity" numbers

TVL and pool-size numbers are easy to fake and easy to misread. A round-trip quote against live state measures the only thing that matters: what you lose entering and exiting right now (fees + price impact, both ways). If a token can be bought but not sold, this tool says so (UNTRADABLE).

Install

npx @iniit/defi-guard-mcp        # or from source: npm install && npm run build

Claude Code

claude mcp add defi-guard -- npx -y @iniit/defi-guard-mcp

Any MCP client (Cursor, Windsurf, etc.)

{
  "mcpServers": {
    "defi-guard": {
      "command": "npx",
      "args": ["-y", "@iniit/defi-guard-mcp"],
      "env": { "BASE_RPC_URL": "https://your-rpc-if-you-have-one" }
    }
  }
}

BASE_RPC_URL is optional; without it the server rotates across public Base endpoints with automatic fallback and retries.

Example

"Before I approve TOKEN X to this router, is any of it risky?"

The agent calls token_safety_screen (can I sell it back? is ownership renounced?), scan_dangerous_capabilities (can the owner mint/blacklist/pause?), and approval_risk (is this an unlimited allowance to an EOA?) — and answers with live on-chain facts instead of vibes, before you sign.

Honesty notes (read this)

Quotes are simulations against live state (eth_call). Real execution adds slippage between quote and inclusion.
Contract addresses (Aave V3 Pool, QuoterV2) are Base mainnet constants, validated against live chain state.
This is a read-only tool. It never holds keys, signs, or submits transactions.
Not financial advice; it reports on-chain state, decisions are yours.

Test

npm run build   # tsc — type-checks and emits dist/. Tools are verified against live Base state.

License

MIT

defi guard mcp