Skip to main content
Version: 2.0.0

CDP's Agentkit Starter Template

Overview​

This template shows an onchain agent powered by Coinbase's AgentKit with the Next.js framework on the frontend and LangGraph for the agent's setup. The agent is designed for AI-driven on-chain capabilities.

AgentKit handles these interactions by using a Gaia node for Large Language Model (LLM) inferencing.

cdp-image

Features​

  • AI-Driven on-chain interactions: Leverages AgentKit to enable AI agents to perform actions on blockchain networks.

  • Bootstrapped: Built as a Next.js project with a LangGraph in the server, bootstrapped with npm create onchain-agent@latest.

  • Configurable LLM: Supports integration with LLMs hosted on Gaia nodes, specifically configured for tool use inferencing (e.g., Llama-3-Groq-8B-Tool, Llama-3.3-70B-Instruct-Q5_K_M).

  • Wallet management: Integrates with a SmartWalletProvider for blockchain interactions, with persistent wallet data management.

  • Extensible actions: Utilizes various Action Providers (e.g., WETH, Pyth, ERC20, CDP API, Wallet actions) to define the agent's capabilities.

  • Chat interface: Provides a user-friendly chat interface for interacting with the agent.

  • Streamed responses: Agent responses are streamed for a more interactive user experience.

  • Memory: Incorporates memory for conversations using MemorySaver from LangGraph.

Getting started​

Prerequisites​

  1. Node.js 18 or later is installed
  2. Confirm that npm 9 or later is installed

Check your Node.js and npm versions:

node --version  # Should be 18+
npm --version   # Should be 9+
  1. You can use either a public Gaia node for example: https://llama70b.gaia.domains/v1 or run the node locally.

Creating a new project​

You can use the CLI to bootsrap a new Agenkit project with the command below:

npm create onchain-agent@latest

Follow the instructions on the CLI to setup your project and choose the Smart wallet (default) option for the setup. You can also make a framework choice among:

  • LangChain
  • Vercel AI SDK

There is also an Model Context Protocol (MCP) option, but, in this guide, we cover using the LangChain option.

Configure secrets and values​

Rename the .env.example to .env and ensure that you have the below values:

CDP_API_KEY_NAME=
CDP_API_KEY_PRIVATE_KEY=

# Optional
NETWORK_ID=base-sepolia

To obtain the values for CDP_API_KEY_NAME and CDP_API_KEY_PRIVATE_KEY head over to CDP portal to create a new API key. Copy the API key name and private key values from the modal that appears.

The NETWORK_ID can stay as base-sepolia and you can explore the possible network options as well.

Project structure​

└── onchain-agent/
    ├── README.md
    ├── next-env.d.ts
    ├── next.config.js
    ├── package.json
    ├── postcss.config.mjs
    ├── tailwind.config.ts
    ├── tsconfig.json
    ├── wallet_data.txt
    ├── .eslintrc.json
    ├── .npmignore
    ├── .yarnrc.yml
    └── app/
        ├── globals.css
        ├── layout.tsx
        ├── page.tsx
        ├── api/
        │   └── agent/
        │       ├── create-agent.ts
        │       ├── prepare-agentkit.ts
        │       └── route.ts
        ├── hooks/
        │   └── useAgent.ts
        └── types/
            └── api.ts

Gaia Integration​

info

A local Gaia node does not require an API key. You will need a Gaia API key to use public nodes.

The LLM inferencing is offloaded to a Gaia node:

  • The LLM is configured in app/api/agent/create-agent.ts.
  • The project uses ChatOpenAI from @langchain/openai to connect to the Gaia node.
  • The specific model configured is "Llama-3-Groq-8B-Tool".
  • The Gaia node endpoint is set via the baseURL in the ChatOpenAI configuration:

For example with a local node running on a machine:

const llm = new ChatOpenAI({
  model: "Llama-3-Groq-8B-Tool",
  configuration: {
    baseURL: "https://YOUR_NODE_ID.gaia.domains/v1", // Gaia node URL
    apiKey: "gaia", // API key for the Gaia node (if required)
  },
});

Run template​

The below command runs the template:

npm run dev

With the template running, there a few example prompts you can use to test the agent:

  • "What is your wallet address?"
  • "What is your wallet balance? Check and confirm."
  • "Share your wallet details including every relevant information."

Documentation​

For further information and advanced topics, refer to the following official documentation: