cache.

cache. is an ERC-4626 vault on Base that turns idle Venice inference credit into compounding yield.

THE PROBLEM

1 DIEM = $1/day of Venice API credit on top-tier and uncensored models, but the credit is daily and does not roll over. Whatever you don't use in 24h is gone. Most holders never use their full allowance, so productive capital evaporates every day it isn't utilised.

THE SOLUTION

Deposit DIEM and the vault stakes it on Venice and resells your idle daily inference credit across neutral marketplaces (Currently on Antseed and Surplus Intelligence). USDC earned is swapped back into DIEM via Aerodrome. Your share, vDIEM, claims its slice of the underlying. An index fund for inference if you will.

HYBRID self-consumption, "have your cache and eat it too": instead of selling all your share, spend part of it yourself as free inference, metered against your vDIEM share. Whatever you don't spend stays in the sell-side path and still earns. This allows users to both benefit from selling or using their inference (without having to commit to just one).

cache. is a vault layer only. It sells THROUGH marketplaces and has no plan to become one. No buyer-side router, no inventory-plus-marketplace conflict.

TECH

• MetaMask Smart Accounts (CORE): the autonomous keeper and seller ARE MetaMask Smart Accounts. Authority is granted via scoped ERC-7715 permissions / ERC-7710 delegations. These are time-boxed so the keeper can compound and nothing else. No admin escape hatch. VERIFIED live on Base: vault.keeper() = the Smart Account 0x12f66323B50F10f51b1D69E1B9FcF1E6668d109A.

• 1Shot: keeper/seller operations are relayed through 1Shot's public relayer with gas paid in USDC in-band. The agent never needs ETH.

• Venice: DIEM is staked on the public Venice contract; the vault's entire yield is real, resold Venice inference. The /inference proxy mirrors Venice's OpenAI-compatible API and advertises ~90 text models.

ARCHITECTURE

CacheVault (ERC-4626) · autonomous keeper/reconciler/dynamic pricer · CacheSellerDelegation (AntSeed seller façade; USDC routes atomically into the immutable vault; capped + timelocked fee switch) · Hono+SIWE+SQLite proxy (meters HYBRID consumption, re-snapshotted every UTC midnight) · CacheUsageLogger (weekly on-chain Merkle attestation, emergency revokePublisher) · Next.js dapp on Vercel.

WHY CACHE?

I built cache for myself first and foremost. I hold DIEM, yet many days I don't use any of the allocated inference. Some days I use it all. Ai / inference marketplaces are already gaining traction (in and out of crypto), and it presents the opportunity to sell the inference you get via DIEM and Venice. I want to sell the inference I don't use, whilst being able to use my inference when I want to. Thus cache.


Proven on Base mainnet: the first gasless harvest, relayed by 1Shot:
tx 0x37e1261de49461f85bc6e31a13170fced311021ae14e0be12c6ae25cf5e77a45

The 1Shot bundler paid the gas and redeemed the scoped ERC-7710 delegation via the DelegationManager, executing harvest() + an in-band USDC fee as the Smart Account.

Going in I had the core ERC-4626 vault concept. Once the core concept was proven, I added the sponsor-aligned systems:

• Re-architected the keeper and seller to operate AS MetaMask Smart Accounts, with authority granted through scoped ERC-7715 permissions / ERC-7710 delegations (compound-only, time-boxed, no admin escape hatch).

• Wired keeper/seller execution through the 1Shot public relayer with gas paid in USDC in-band. Useful as we do not earn ETH and thus a cost in ETH doesn't make sense (even if small on Base)

• Built CacheSellerDelegation, an AntSeed seller facade that routes USDC atomically into the immutable vault,

• Built HYBRID self-consumption: a Hono + SIWE + SQLite proxy that meters each holder's usable inference against their vDIEM share, billed atomically and re-snapshotted every UTC midnight.

• Built CacheUsageLogger for weekly on-chain Merkle attestation of usage, with emergency revokePublisher. This is mainly to 'prove' the ussage, and that the protocol (or others) is not using depositors inference.

• Shipped the /inference dapp page (AES-GCM-encrypted API key in localStorage) mirroring Venice's OpenAI-compatible API across their models.

• Deployed all contracts to Base mainnet (Safe-owned) and VERIFIED the Smart Account keeper live: vault.keeper() = 0x12f66323B50F10f51b1D69E1B9FcF1E6668d109A, seller.isOperator(SA) = true.

• Hardened to 164 forge + 54 proxy + 105 scripts tests, Slither 0 high / 0 medium.

No fundraising. There is no $CACHE token. Self-funded, open source, contracts Safe-owned with no admin escape hatches.
The DIEM currently in the vault (1 DIEM from 1 depositor) was purchased to put DIEM in the vault / skin in the game.