More details on what Matt mentioned

What We Built

MCP server + MCP UI toolkit that adds wallet management and treasury operations to AI agents. Works in:

• OpenClaw (Moltbot, Clawdbot, etc.)
• Claude Desktop / Claude Code
• Other MCP-compatible agent frameworks

The MCP UI renders interactive approval cards directly in the agent's interface — transaction details, approve/reject buttons, multi-sig flows. Users see exactly what the agent wants to do and approve with a tap.

With OpenClaw exploding right now (1.5M+ agents on Moltbook, agents tipping each other, wild takes from Elon to Karpathy), the timing feels right to show what safe agent finance actually looks like.

How Turnkey Fits

Sub-org per customer, shared quorum:

• Root User 1: Customer passkey (their device)
• Root User 2: Catena API key (our treasury agent)
• Threshold: 2 — neither signs alone

Policies define what the agent can do automatically:

// Agent can deposit to whitelisted vaults alone
{
  "effect": "EFFECT_ALLOW",
  "consensus": "approvers.any(user, user.id == '<CATENA_AGENT>')",
  "condition": "eth.tx.to in ['0xMorpho...', '0xAave...'] && eth.tx.value < 10000000000000000000"
}

// Withdrawals need customer co-sign
{
  "effect": "EFFECT_ALLOW",
  "consensus": "approvers.count() >= 2",
  "condition": "eth.tx.function_name == 'transfer'"
}

The key: policies run in Turnkey's Nitro enclave, not our app code. Even if our backend is compromised, the enclave won't sign transactions that violate customer-defined rules.

Architecture