Solana News Today: SendAI Releases Guide to Customize Solana Agent Kit with New On-Chain Functions

SendAI has released a detailed guide on customizing the Solana Agent Kit, an open-source framework for building Solana-native autonomous agents that interact with Solana programs and off-chain APIs through natural language prompts. The guide outlines how developers can expand the toolkit’s capabilities by adding new on-chain functions, such as fetching a liquidity pool’s trade history from CoinGecko [1]. Developers are guided through the process of cloning the repository, setting up prerequisites like a CoinGecko API key, and configuring the development environment with node.js and pnpm [1].

The Solana Agent Kit operates by integrating with large language models (LLMs) such as OpenAI and Perplexity. These models support a feature known as function calling, where the AI agent suggests the use of specific tools based on the input it receives. However, the actual execution of the function is managed by the Solana Agent Kit, which interprets the AI’s suggestion and executes the corresponding function [1]. This workflow is streamlined using LangChain, which handles the parsing of function calls and the handoff of results back to the LLM, simplifying the interaction between the model and external logic [1].

Customization involves defining new tools in a JSON specification and creating the associated handler functions that execute when the agent interprets the function call. For instance, to fetch a pool’s trade history, a new file `get_latest_trades_by_pool.ts` is created under the `src/tools/coingecko` directory. The function is exported and integrated into the main agent class in `src/agent/index.ts`, allowing it to be recognized and used by the agent [1]. Additionally, a corresponding LangChain tool is defined in `src/langchain/coingecko/get_latest_trades_by_pool.ts`, specifying the function’s name, input, output, and the underlying logic [1].

To ensure the new functionality is accessible to the language model, an action must be defined in `src/actions/coingecko/getCoingeckoLatestTradesByPool.ts`. This action includes a description, example inputs and outputs, and a handler function. Once defined, the action is registered in `src/actions/index.ts`, making it available to the agent for execution [1].

The guide also includes testing procedures, such as setting up a `.env` file with required credentials, installing dependencies with pnpm, and running the application with `pnpm run test`. This allows developers to verify that their customized agent can perform the new functions using natural language prompts [1].

Despite its powerful capabilities, the Solana Agent Kit is still in the experimental phase. Developers should be aware that LLMs can occasionally produce incorrect or fabricated outputs, which may lead to errors or unintended actions. Additionally, the agent may select the wrong function if the task description is not clear enough, highlighting the importance of precise action definitions [1]. Given the experimental nature of the technology, it is advisable to avoid using it with real wallets or production environments until its reliability is confirmed [1].

The guide encourages further exploration by directing readers to additional resources, such as articles on building AI models for crypto price prediction, suggesting that the platform is actively evolving with new features and applications [1].

Source: [1] How to Customize the Solana Agent Kit by SendAI

https://www.coingecko.com/learn/customize-solana-agent-kit