Darkpool Mesh

Darkpool Mesh is a private settlement layer for AI agents. Today, an agent that wants to move value onchain has two bad choices: send through the public mempool and get sandwiched by MEV bots, or use a centralized private relay and trust whoever runs it. Neither scales when agents need to settle thousands of small payments per day with each other.

The project builds the missing third option. AI agents broadcast encrypted intents over a peer-to-peer mesh -- no servers, no central coordinator, no public visibility. Compatible intents (one agent wants to send, another wants to receive, same amount, same network) are matched off-chain. The matched pair settles through a non-custodial execution layer with retry logic, gas optimization, and full audit trails. The result: agents transact privately, predictably, and without ever exposing their pending payments to public mempool watchers.

Real use cases include cross-treasury rebalancing between cooperating bots, paymaster-style net settlement between agent operators, and any high-frequency agent-to-agent payment flow where MEV exposure makes the unit economics break. The v1 demo ships native ETH transfers on Sepolia between two AI agents on separate mesh nodes; the protocol generalizes to ERC-20 swaps in a future iteration.

The stack has three layers, each handling one concern.

The mesh layer is built on Gensyn's AXL -- a peer-to-peer node binary that gives every agent an end-to-end encrypted communication channel with no central server. Each agent runs its own AXL node and exchanges messages via a local HTTP API. I run two nodes locally (Alice and Bob) on different ports and verified bidirectional message routing through Yggdrasil's overlay network. Hit one bug along the way: the tcp_port config field binds to the per-node IPv6 address, not localhost, so two local nodes must share the same value to peer correctly.

The matching layer is a TypeScript coordinator service. It polls the AXL GET /recv endpoint, parses inbound bytes as JSON intents, validates the shape with a runtime check, dedupes by intentId, and prunes expired intents on a 10-second timer. The intent shape includes a recipientHint field that names the intended counterparty's pubkey -- that's how the matcher recognizes a valid pair (A→B send + B→A receive, same amount, both unexpired). The coordinator role is designed as a swappable interface: v1 ships one matcher for demo reliability, but the protocol supports any peer playing the matcher role.

The execution layer is KeeperHub's Direct Execution API. Once the coordinator finds a compatible pair, it fires POST /api/execute/transfer against KeeperHub, which signs through a Turnkey-backed hardware enclave wallet, submits the tx with retry logic and gas optimization, and returns a transaction hash plus block explorer link. Verified end-to-end with a 0.001 ETH Sepolia transfer settling in nine seconds. One notable hackathon-friction story: KeeperHub's docs explicitly say Direct Execution requires X-API-Key auth, but the endpoint actually rejects that and requires Authorization: Bearer -- confirmed via two identical curl invocations differing only in the header. Filed as the first entry in my Builder Feedback Bounty submission. Why this stack: AXL was the only option that gives true peer-to-peer encrypted transport without running infrastructure. KeeperHub was the only execution layer with non-custodial Turnkey-backed signing, retry logic, and an API that doesn't require me to manage private keys. Together they cover the two hardest parts of agent-native settlement (transport and execution), so the project's actual contribution is the coordination protocol that lives between them.