Loops HouseNEAR Sentinel: Did It Land?

NEAR HACKATHON
Links
Team
1 member- KAOwner
Karios
Overview
The RPC timed out. Did the payment land? An AI agent that sends a NEAR payment and gets a timeout can't tell whether the money moved. Sign again and it may pay twice; don't, and it may never pay.
NEAR Sentinel answers with proof before the agent signs anything new. It asks several independent RPC providers and returns: landed (DO_NOT_RETRY), can still land (WAIT_AND_RECHECK: re-send the same bytes only), can never land (SAFE_TO_RETRY, proven because its nonce was used or its validity window expired, on at least two independent operators), or partly executed (MANUAL_REVIEW_REQUIRED: a top-level success hid a failed downstream call). Unknown is never reported as safe.
A guard contract only knows about calls that reach it; if the call to the guard itself times out, only the chain can say whether it will land. Sentinel answers that for any NEAR transaction, guarded or not, with no contract to adopt, using what NEAR makes provable: the access key's nonce, the validity window and the receipt model, checked across independent RPC operators. It also works read-only on mainnet.
Proven on testnet with real money: the naive agent paid twice, the guarded agent once; an unsent payment was answered WAIT, then proven SAFE once its nonce was used.
For agents: an MCP server (also an Agent Skill, designed to sit next to NEAR's official MCP Server), a TypeScript SDK and a REST API; an example agent on NEAR AI Cloud (TEE-hosted model) where Sentinel's verdict is binding. On-chain: Sentinel Guard (pay at most once per intent) and Sentinel Vault (an allowance an agent spends with a function-call key; owners manage it via NEAR Connect). Live dashboard with one-click "Run it live": https://64-177-60-105.sslip.io. One command verifies everything: pnpm verify.