The problem Sidekick solves

Sidekick is a fully autonomous agent that plays onchain games. Players use it automate in-game tasks - gathering resources, protecting their base etc. Given a contract address and a goal, it learns how to play the game and transacts on behalf of the player, without any input.

User Interaction and Data Flow

Sidekick runs as a widget in the browser. Users type in a goal for agent to execute then hit "play", and it will continously play the game without any further input from the player. The model's transactions and reasoning are displayed in the widget.

The project architecture and development process

Sidekick runs in the browser. Given a smart contract address, we fetch the contracts source code and current storage values. This is used to prompt Claude, via Langchain to generate a transaction to accomplish the users goal. Once the transaction is generated, Sidekick executes it onchain. Once the transaction is confirmed, Claude fetches the blockchain state again and restarts the loop.

Product Integrations

Sidekick uses Claude for the LLM. The example game is deployed on Base.

Key differentiators and uniqueness of the project

At it's core, Sidekick is a fully autonomous agent. Given a contract address and goal, it can reason and transact on behalf of the player without any further input.

Importantly though, Sidekick runs in the browser, so you can use while playing onchain games. You can stop and start Sidekick at will, simply ask it for advice, or just stop it entirely and play the game yourself.

Trade-offs and shortcuts while building

The logic for passing the Sidekick the contract source code is not generic now, it requires running a script to generate Typescript files that the model can read. In the future I would like it to read automatically or from a source like Etherscan.

Additional Features

I started this project during the hackathon