This guide is accompanied by a video walkthrough.

Before diving into specific tools, let's understand the key components in the AI development stack:

Understanding the AI Stack

1. Large Language Models (LLMs)

   • The foundational AI models (like GPT-4, Claude, or Llama)
   • Process and generate natural language
   • Provide raw intelligence and reasoning capabilities
   • Examples: OpenAI's GPT-4, Anthropic's Claude, Meta's Llama

2. Agent Frameworks (like Eliza)

   • Built on top of LLMs to create autonomous agents
   • Manage agent personalities and behaviors
   • Handle memory and context
   • Provide multi-platform integration
   • Examples: Eliza, LangChain, AutoGPT, Vercel AI

3. Domain-Specific Agent Kits (like Solana Agent Kit)

   • Built for specific use cases (e.g., blockchain operations)
   • Provide pre-built tools and integrations
   • Abstract away technical complexity
   • Ready-to-use blockchain functions
   • Examples: Solana Agent Kit, GOAT Toolkit

Think of it like this:

• LLMs are the "brain" that can understand and generate language
• Agent frameworks are the "operating system" that manages the AI's behavior
• Domain-specific kits are the "apps" that give the AI specific capabilities

Building AI Agents for Solana

This guide will show you how to:

• Build a basic AI agent using SendAI's Solana Agent Kit
• Create a Twitter bot with Eliza Framework
• Understand on-chain AI integration options
• Options on how to build advanced trading bots

Introduction to AI Agents on Solana

AI agents on Solana are autonomous programs that can interact with the blockchain using natural language processing and machine learning. They can:

• Execute transactions
• Monitor blockchain activity
• Interact with smart contracts
• Provide user assistance
• Automate complex workflows

Available Frameworks

Several frameworks are available for building AI agents on Solana:

1. SendAI's Solana Agent Kit

• A collection of tools for building agents on Solana
• Compatible with Eliza, Langchain, and Vercel AI SDK
• Built-in Solana tools and utilities
• Natural language processing capabilities
• Solana Agent Kit
• Documentation

2. Eliza Framework

• Combines datalayer, llm and agents and has a dedicated Solana plugin
• Extensible agent architecture
• Easy integration with existing projects and can be used for example for Twitter (X), telegram or discord bots
• GitHub Repository

3. GOAT Toolkit

• Comprehensive development toolkit
• Cross-chain compatibility
• GitHub Repository

4. Rig Framework

• Native Rust implementation
• Good for high performance trading bots
• Can be combined with the Listen.rs toolkit for Solana integration transaction monitoring and sending transactions using Jito bundles for example
• GitHub Repository
• Listen.rs Repository

If you are building an AI framework feel free to open a PR to add your framework to the list

Getting Started with SendAI

Let's start by setting up a basic AI agent using SendAI's Solana Agent Kit that can perform actions on the Solana blockchain and that we can chat with:

node --version  # Check current version
nvm install 23  # Install Node.js 23 if needed
nvm use 23      # Switch to Node.js 23

pnpm install solana-agent-kit
pnpm add @langchain/core @langchain/openai @langchain/langgraph dotenv bs58

{
  "dependencies": {
    "@langchain/core": "^0.3.33",
    "@langchain/langgraph": "^0.2.41",
    "@langchain/openai": "^0.3.17",
    "bs58": "^6.0.0",
    "dotenv": "^16.4.7",
    "solana-agent-kit": "^1.4.3"
  }
}

OPENAI_API_KEY=your_openai_api_key
RPC_URL=https://api.devnet.solana.com
SOLANA_PRIVATE_KEY=your_private_key

solana-keygen grind --starts-with ai:1

import { SolanaAgentKit, createSolanaTools } from "solana-agent-kit";
import { HumanMessage } from "@langchain/core/messages";
import { ChatOpenAI } from "@langchain/openai";
import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { MemorySaver } from "@langchain/langgraph";
import * as dotenv from "dotenv";
import bs58 from "bs58";
import * as readline from "readline";
 
dotenv.config();
 
async function initializeAgent() {
  const llm = new ChatOpenAI({
    modelName: "gpt-4-turbo-preview",
    temperature: 0.7,
  });
 
  // Convert array string to actual array, then to Uint8Array, then to base58
  const privateKeyArray = JSON.parse(process.env.SOLANA_PRIVATE_KEY!);
  const privateKeyUint8 = new Uint8Array(privateKeyArray);
  const privateKeyBase58 = bs58.encode(privateKeyUint8);
 
  const solanaKit = new SolanaAgentKit(privateKeyBase58, process.env.RPC_URL!, {
    OPENAI_API_KEY: process.env.OPENAI_API_KEY!,
  });
 
  const tools = createSolanaTools(solanaKit);
  const memory = new MemorySaver();
 
  return createReactAgent({
    llm,
    tools,
    checkpointSaver: memory,
  });
}
 
async function runInteractiveChat() {
  const agent = await initializeAgent();
  const config = { configurable: { thread_id: "Solana Agent Kit!" } };
  const rl = readline.createInterface({
    input: process.stdin,
    output: process.stdout,
  });
 
  // Clear console and start chat with a small delay
  setTimeout(() => {
    console.clear(); // Clear any initialization messages
    console.log("Chat with Solana Agent (type 'exit' to quit)");
    console.log("--------------------------------------------");
    askQuestion();
  }, 100);
 
  const askQuestion = () => {
    rl.question("You: ", async (input) => {
      if (input.toLowerCase() === "exit") {
        rl.close();
        return;
      }
 
      const stream = await agent.stream(
        {
          messages: [new HumanMessage(input)],
        },
        config,
      );
 
      process.stdout.write("Agent: ");
      for await (const chunk of stream) {
        if ("agent" in chunk) {
          process.stdout.write(chunk.agent.messages[0].content);
        } else if ("tools" in chunk) {
          process.stdout.write(chunk.tools.messages[0].content);
        }
      }
      console.log("\n--------------------------------------------");
 
      askQuestion(); // Continue the conversation
    });
  };
}
 
runInteractiveChat().catch(console.error);

npx tsx agent.ts

Please show me my wallet address and request some devnet sol

Please create me a NFT collection called trains with symbol TRN using this uri: https://scarlet-fancy-minnow-617.mypinata.cloud/ipfs/bafkreif43sp62yuy3sznrvqesk23tfnhpdck4npqowdwrhrzhsrgf5ao2e

Please mint me an NFT into that collection using the name: Train1 and using this URI: https://scarlet-fancy-minnow-617.mypinata.cloud/ipfs/bafkreif43sp62yuy3sznrvqesk23tfnhpdck4npqowdwrhrzhsrgf5ao2e

Where to go from here?

• You can now for example import the private key into your browser extension wallet to see the NFT.
• You can ask the agent to show you all your NFTs. You will notice you will get an error for this action. This is because the default action to request assets uses the Helius Asset api to request assets so for that you would need to add a Helius API key to your .env file and pass it into the agent. pass it into the agent.

const solanaKit = new SolanaAgentKit(privateKeyBase58, process.env.RPC_URL!, {
  OPENAI_API_KEY: process.env.OPENAI_API_KEY!,
  HELIUS_API_KEY: process.env.HELIUS_API_KEY!,
});

• You can now start writing your own actions following the Contribution guide

There is also an example of a web chat for your agent using React that you can for example deploy to Vercel.

Not all actions are working 100% yet, like the request balance for example sometimes fails, but you can see the progress in the Solana agent kit repo and also contribute your own actions to be used by everyone.

Building with Eliza Framework

The ElizaOS (formerly known as Ai16z) is an AI framework. That means it combines multiple applications into one and organizes their interactions. It combines:

1. Different LLMs to understand the user
2. Blockchain agents with different actions (like minting an NFT for example)
3. Datalayers to store the characters state between actions using for example Redis, PostgreSQL DB or local storage
4. Different clients like X, telegram, discord. etc.

Its takes all these things and puts them together in a way that you can create a "Character" that can interact with users or autonomously. These Characters are defined in a JSON file, which defines which actions it can perform, which LLM it should use and different prompts that define the behaviour of the character.

Building a Twitter (X) bot that can mint NFTs on request

What we will build now is a helpful solana developer assistant that can help developers on the platform X to get started with solana and it will also be able to mint NFTs to the users on demand.

git clone https://github.com/elizaOS/eliza.git

node --version  # Check current version
nvm install 23  # Install Node.js 23 if needed
nvm use 23      # Switch to Node.js 23

pnpm install --no-frozen-lockfile
pnpm build

cp .env.example .env

TWITTER_USERNAME=               # Account username
TWITTER_PASSWORD=               # Account password
TWITTER_EMAIL=
OPENAI_API_KEY=                 # Get this from: https://platform.openai.com/api-keys
SOLANA_RPC_URL=https://api.devnet.solana.com # Lets work on devnet for now
SOLANA_PRIVATE_KEY=             # Get this from: https://solana.com/wallet
SOLANA_PUBLIC_KEY=
SOLANA_CLUSTER=devnet # Default: devnet. Solana Cluster: 'devnet' | 'testnet' | 'mainnet-beta'
SOLANA_ADMIN_PRIVATE_KEY=       # This wallet is used to verify NFTs (you can use the same as the SOLANA_PRIVATE_KEY for now)
SOLANA_ADMIN_PUBLIC_KEY=        # This wallet is used to verify NFTs (you can use the same as the SOLANA_PUBLIC_KEY for now)

const bs58 = require('bs58');
console.log(bs58.encode(new Uint8Array([213,29,143,...])));

npx tsx encode58.ts

{
    "name": "Sol",
    "clients": ["twitter"],
    "modelProvider": "openai",
    "settings": {
        "voice": {
            "model": "en_US-neural"
        },
        ...

pnpm start --character="characters/sol.character.json"

pnpm start --character="characters/sol.character.json" --host
pnpm start:client

Disclaimer: The Solana integration and actions are not perfect yet but they are in constant development and are moving fast. There will be a direct integration of the SendAI agent kit into Eliza soon.

Using GOAT Toolkit

GOAT is similar to Solana Agent Kit a collection of many different blockchain tools that also supports different Solana actions and other chains. It can also be integrated into the Eliza framework for example and supports many different wallets, frameworks and block chains.

On chain program agent integration

If you want to use an agents reply within a solana program you need to use an oracle that writes the reply data into an account that your program can read. You can find an example of an LLM which will transfer tokens to any user that answers questions correctly here:

On chain program LLM interaction

This is an example of an on-chain program that integrates with an LLM (AI) oracle to create an interactive token-dispensing agent named "Mar1o".

Key Features

1. AI Agent Interaction

   • Users can chat with Mar1o to try to earn tokens
   • Agent evaluates user's Solana knowledge to decide token amounts
   • Responses are in JSON format with reply text and token amount

2. Token Management

   • Creates MAR1O tokens with metadata
   • Can mint between 0-10,000 tokens based on AI response
   • Uses PDAs for token management

Program flow explained

Here's how this program works:

// Initialize the AI agent with personality
const AGENT_DESC: &str = "You are an AI agent called Mar1o which can dispense MAR1O tokens...";
 
// User interaction flow:
// 1. User sends message to agent
// 2. Oracle processes with LLM
// 3. Agent responds with JSON containing reply and amount
// 4. If amount > 0, tokens are minted to user
 
// Example response:
{
    "reply": "Great answer! Here's a reward.",
    "amount": 1000
}

This demonstrates how to combine AI capabilities with on-chain token mechanics to create interactive blockchain experiences. Here is a live example where you can try to convince the Agent to transfer you some tokens.

Advanced Rust agents for algorithmic trading

If you want to write your agents in Rust you can use the Rig framework and the Listen.rs library.

These two work nicely together. While the Rig framework gives you an easy way to connect to LLMs the Listen.rs toolkit provides you the Solana integration including Transaction monitoring and sending transactions using Jito bundles for example.

Best Practices

• Know that your agent has a private key and thus can access the funds in your wallet. If someone can chat with your agent and it has a transfer action for example users can probably get it to transfer them your funds.
• Use a payed RPC endpoints. The public endpoints usually get rate limited
• Monitor agent activities and make sure they are not doing anything malicious or get tricked into it by users interacting with them
• Implement priority fees and transaction retries to your agent actions so that they also land when there is congestion on the network
• If possible test on devnet first before you deploy to mainnet. Most agents just take the RPC url as an input and you can use a devnet one until your are ready for mainnet. Not all actions maybe possible to do on devnet though.
• When writing your own actions add integration tests for them so that the frameworks can monitor which actions are failing due to changed programs or APIs

Conclusion

AI agents on Solana open up new possibilities for automation and interaction with the blockchain. While not all integrations are perfect yet they are developed fast and offer lots of opportunity for contributions and are very flexible of adding your own integrations.