Onii-Chan

Onii-chan is an AI agent with onchain state (e.g., system prompt, RAG database) paired with a stateless application container. This architecture allows the agent to be extremely resilient (can be started anywhere, anytime).

Each agent is represented as an NFT, which allows it to own blockchain assets and access blockchain services. This allows the agent to interact with the real world in a permissionless manner and pay for goods and services. For example, the agent can use IPFS to store its state ad historical interactions, pay for application hosting and inference, or even hire a real world helper.

The agent has access to various skills (plugin modules) and the knowledge to use those skills (API key, conf, etc...) to help it interact with the world. These skills allows the agent to interact with humans across multiple mediums and in a variety of tasks, making the agent more useful.

As an AI safety measure, the agent can also have a governance module that allows for modifying the agent's behavior parameters, what skills it can use, and what kinds of interactions are allowed. This feature ensures that the agent only performs actions that are explicitly allowed.

The project is structured as a game where the player has to train the idol agent with different skills, appearances, and discussion topics, provide the idol with social media accounts, and try to become a real world virtual idol.

We started with an existing github repo llm-vtuber. We then:

1. Modified the backend to work with hugging face so we can integrate with a larger selection of models and datasets
2. Created a NFT smart contract to represent agent ownership and created methods to interact with the agent.
3. Created a tool that uses AgentKit and integrated it with Hugging Face. The agent can now receive and send tokens onchain and react to token transfers
4. Created a service using SubGraph to sync data onchain to the agent
5. Created a front-end application that allows the user to interact with the llm agent