Their AI uses the blockchain.Our AI learns from it.

We train models directly on transaction data to build the first blockchain-native AI systems that actually understand the environment they operate in.

Let's zoom in on why that matters…

Today's on-chain agents are powered by large language models. But blockchain activity isn't recorded as words and sentences.

Drop these models on-chain and the most powerful reasoning engines ever built are surrounded by data they never learned to understand.

We're here to change that.

Our first models are already live

Different products. Different models. Same fundamental advantage: they learn directly from blockchain data.

.01For Users

Hōkū

Hōkū is the trading app we built on a suite of over 40 models. It watches clusters of wallets in real time and flags a trend while it's forming, not after it's on a chart. Live now on Ethereum, Base, and BNB Chain.

Try it now
.02For Agents

MCP

Agents today react to charts and social feeds, which means they act on information the market already has. Our MCP server gives an agent direct access to what's happening next on-chain, in a format it can understand.

.03For Builders

Models

The same models are available directly, deployable in a line of code. Predictive signals, advanced advertising and proactive security, all trained on transaction data, built for teams running their own systems.

Learn more

Training data that never stops growing

The blockchain writes its own training data, every second of every day. Each new block adds new behaviors for our models to learn from.

Transactions used in training
12,118,290,585
Wallets used in training
641,250,259
Blocks ingested in inference
194,856,446
Events used in training
41,624,751,616
419,038trade signals generated with
90%Accuracy
90%

AI learns from the chain. Here's the proof.

Our models learned patterns in on-chain trading behavior. Now they find them for you live, before the markets move.

Soon, the first large blockchain model

A large blockchain model starts with the same principle that gave LLMs their extraordinary capabilities: neural networks learning from vast sequences of data. For LLMs, those sequences are sentences. On-chain, they're the trillion actions already ordered inside blocks. Learning from those sequences reveals behaviors, patterns and context on blockchains at a scale no human could ever identify.

One model, endless applications

Prototype models already show measurable learning that will scale up to frontier-grade intelligence for every business, user and agent on Web3. Here's a deeper look inside how it works.

journalctl-fu encoder
tail-f selectors.loglive
  1. 010xa9059cbb ETHtransfer(address,uint256)
  2. 020x3593564c BASEexecute(bytes,bytes[],uint256)
  3. 030x791ac947 BSCswapExactTokensForETHSupportingFeeOnTransferTokens(uint256,uint256,address[],address,uint256)
  4. 040x095ea7b3 ETHapprove(address,uint256)
  5. 050xac9650d8 BASEmulticall(bytes[])
  6. 060x7ff36ab5 BSCswapExactETHForTokens(uint256,address[],address,uint256)
  7. 070x617ba037 ETHsupply(address,uint256,address,uint16)
  8. 080xd0e30db0 BASEdeposit()
z.npyfloat32[768]
tail-f decoded.logregion of z

The encoder emits a 768-dimension behavioural state. Signed vector values emerge from the neural network, gather, and transition into a decoded behaviour description. This illustrates a behavioural representation, not a prediction or confidence score.

Built by alumni from

  • Google
  • Apple
  • Facebook
  • Fidelity Investments
  • Massachusetts Institute of Technology
  • Stanford University
  • The Wharton School