@lombard.finance/sdk-agentkit

Coinbase AgentKit ActionProvider for the Lombard protocol. Enables AI agents built with AgentKit to stake Bitcoin, mint LBTC, check balances, and deploy to DeFi vaults.

Installation

npm install @lombard.finance/sdk-agentkit
# or
yarn add @lombard.finance/sdk-agentkit

Peer Dependencies

npm install @coinbase/agentkit viem zod

Quick Start

import { lombardActionProvider } from "@lombard.finance/sdk-agentkit";
import { AgentKit } from "@coinbase/agentkit";

const agentkit = await AgentKit.from({
  walletProvider,
  actionProviders: [
    lombardActionProvider({
      confirmWrite: (request) => askTheOperator(request),
    }),
  ],
});

// Agent now has access to all Lombard actions
const actions = agentkit.getActions();

Approving Writes

A tool call is the transaction: the write actions sign and send as soon as they are invoked, and what invokes them is a model reading text. Not all of that text is yours — a token symbol, an address label or an error relayed from an upstream service reaches the model too, and a tool call cannot be distinguished from an instruction after the fact.

So each write asks first. Give the provider exactly one of:

• confirmWrite(request) — called before anything is signed, including the fee authorisation some chains require. Return false, or throw, and nothing is sent. request carries the action, chain, account, amount and assets, plus recipient when the funds land somewhere other than the signing account. That is the field to read twice: on the native BTC output route it is a Bitcoin address supplied as a tool argument. It is absent where the transaction pays the signer, so a prompt never names a destination the transaction does not use.
• autoApproveWrites: true — no confirmation at all, for a wallet that is meant to run unattended and is funded accordingly.

With neither set, write actions report that confirmation is unconfigured and sign nothing. Read actions are never gated.

The two are mutually exclusive and passing both throws at construction. They contradict each other, and the way to end up with both is adding confirmWrite to a config that already carried autoApproveWrites: true — connecting an approval flow and leaving the old flag behind. Picking a winner silently would mean handing back unattended execution to someone who thinks they just built a gate, so it fails while the config is still in front of whoever wrote it.

Asking the model to confirm in its system prompt is worth doing and is not the same thing: that is a request to the model, this is a gate it cannot talk its way past.

With LangChain

import { lombardActionProvider } from "@lombard.finance/sdk-agentkit";
import { AgentKit } from "@coinbase/agentkit";
import { getLangChainTools } from "@coinbase/agentkit-langchain";
import { createReactAgent } from "@langchain/langgraph/prebuilt";

const agentkit = await AgentKit.from({
  walletProvider,
  actionProviders: [
    lombardActionProvider({ confirmWrite: (request) => askTheOperator(request) }),
  ],
});

const tools = getLangChainTools(agentkit);
const agent = createReactAgent({ llm: yourModel, tools });

Available Actions

Write Actions

| Action | Description | | ------ | ----------- | | stake_btcb_to_lbtc | Stake BTC.b to receive LBTC. Handles approval and fee authorization automatically | | unstake_lbtc_to_btc | Unstake LBTC to native BTC (cross-chain) or BTC.b (same-chain) | | redeem_lbtc_to_btcb | Simple same-chain LBTC to BTC.b conversion | | deploy_to_earn | Deploy LBTC into a DeFi vault (Bitcoin Earn) for additional yield | | claim_lbtc_deposit | Claim a notarized BTC deposit to mint LBTC |

Read Actions

| Action | Description | | ------ | ----------- | | get_lbtc_balance | Check LBTC balance for any address on the current chain | | get_btcb_balance | Check BTC.b balance for any address on the current chain | | get_lbtc_exchange_rate | Current LBTC/BTC exchange rate and minimum stake amount | | get_deposit_status | Track all deposits (pending, claimable, claimed, failed) | | get_redemption_status | Track all unstake and redemption operations |

Supported Networks

| AgentKit Network ID | Chain | Environment | | ------------------- | ----- | ----------- | | ethereum-mainnet | Ethereum | Production | | ethereum-sepolia | Ethereum Sepolia | Testnet | | base-mainnet | Base | Production | | base-sepolia | Base Sepolia | Testnet |

Network aliases are also accepted: ethereum, eth, mainnet, sepolia, base, base-sep.

How It Works

The LombardActionProvider bridges Coinbase AgentKit and the Lombard SDK:

1. Network mapping: Translates AgentKit network IDs to Lombard chain configuration
2. Wallet adaptation: Converts AgentKit's wallet provider to EIP-1193 for the Lombard SDK
3. Automatic approvals: Checks and sets token allowances before transactions
4. Fee authorization: Handles EIP-712 fee signatures on chains that require them (Ethereum, Sepolia)
5. Error sanitization: Strips RPC URLs and sensitive data from error messages

All write actions execute real on-chain transactions through the AgentKit wallet.

Utilities

import {
  resolveNetwork,
  resolveChainName,
  isLombardSupportedNetwork,
} from "@lombard.finance/sdk-agentkit";

// Check if a network is supported
isLombardSupportedNetwork(network); // boolean

// Resolve AgentKit network to Lombard config
const resolved = resolveNetwork(network);
// { chainId, env, networkId }

// Resolve by friendly name
const resolved = resolveChainName("ethereum");
// { chainId: 1, env: "prod", networkId: "ethereum-mainnet" }

Requirements

• Node.js: 18+
• @coinbase/agentkit: 0.2+
• viem: 2.21+

License

MIT