The protocol is dual-licensed:
• AGPL-3.0 for the core protocol implementation
• MIT for interfaces, libraries, and integration tooling
Inspect it.
Audit it.
Fork it.
Build on top of it.
The next generation of digital-asset protection infrastructure will be open-source by default.
@BaseHubHB@baselatam@EFetheverywhere@CoinbaseDev@ethereum
Mr Haven’s smart contracts are now public.
The code securing live and programmable time vaults on @base is open:
verifiable, auditable, and inspectable by anyone.
https://t.co/77Ibi7SXmH
🧵
Trillions of dollars are moving onchain.
But the systems used for long-term wealth coordination — estate planning, inheritance distribution, beneficiary management, insurance coverage — were designed for traditional financial rails.
They do not map cleanly to digital-native assets, global families, or self-custody.
At the same time, self-custody without recovery introduces its own failure mode.
Onchain infrastructure can solve both problems.
But only if the underlying primitives are open.
T4
Mr Haven V1 has been live on Base since December 2025:
https://t.co/EykATO3e8h
Core contracts:
VaultRouter
0x8C09a5f79e62D29BeB9d2FA710Ecc4cC17cdE842
PlanExecutor
0x244eA67F2FcaEBeEBeCdFF1cE45a9abB4918E02D
EmailEscrowManager
0x2E3E2aC4F32A0eEe3b811D302fBB897f1E027b25
Verify the deployments on Basescan.
Recompile from the v1.0.2 release tag.
Bytecode is the proof.
T5
Each user receives a dedicated personal vault.
While assets remain time-locked, capital can earn yield through Aave V3.
Distribution logic is governed by user-defined plans:
• scheduled execution
• inactivity-triggered execution
• beneficiary-directed distribution
Beneficiaries can be specified either by wallet address or by email using hash-locked escrow flows.
For the past five months, I’ve worked 70+ hour weeks to bring this idea to life. The thesis is simple: if AI is going to excel at computer use, it should be the computer itself.
Today a milestone. Still early, much more ahead. 👇
So I re-ran with measurement. Of the ~80s wall: 0s local LLM inference, ~55s waiting on Delta (server roundtrips + dwell we program in to look human), ~27s in the skill substrate (verification + browser snapshots).
The model isn't in the loop on the happy path. The skill is a predefined recipe for this site and others like it. Local Qwen 27B steps in only when a step fails its verification gate.
If you have a recurring workflow you wouldn't trust to a normal cloud agent, healthcare, admin, legal, compliance, back-office, let's talk.
▎https://t.co/6tXC1LMQV3
This isn’t “small model beats GPT-5.5.”
It’s “encode the workflow, and dispatch becomes the bottleneck.”
What’s still hard: anti-bot reputation, cross-site transfer, and durability over days.