presign-guard for NEAR

AI agents on NEAR can be talked into signing things that drain them: a full-access key added to their account, the account deleted, code deployed over the wallet, a key added to NEAR Intents, a swap that gives back a fraction of its value, or a deposit into a look-alike of intents.near. presign-guard is a pre-sign check that already protects agents on EVM, Solana, XRPL, Algorand and Stellar; for this hackathon we built its NEAR support. Before an agent signs, it sends the NEAR transaction (borsh or wallet-selector JSON, meta transactions included) or the NEAR Intents message (NEP-413, also inside execute_intents) and gets green, orange or red with typed reasons. It reads the account's access keys and token metadata over NEAR RPC, values Intents swaps at NEAR Intents' own 1Click prices, flags fake USDC/USDT/wNEAR and look-alike accounts, and screens every party against the OFAC SDN list. The function-call key is NEAR's sandbox for an agent; presign-guard tells the agent when a transaction would break out of it. Demo (node near/demo.mjs): a real 1Click quote, six cases, the honest swap and a scoped function-call key signed, a tampered quote, an injected key, a look-alike deposit and a prompt-injected full-access key refused. Free verdict endpoint and MCP tools live at presign-guard.fizzl.eu.