Make AI agent indexable and verifiable on the blockchain.

Version: 0.2

Last Updated: 2024-06-15


What is AI721

AI721 introduces a standard that extends the AI capabilities of ERC721. It's used to index AI agent on chain, describe the AI personality and copabilities of your AI agent, and verify if a specific personality belongs to a particular AI agent. This makes your AI agent become digital asset on the blockchain.

AI721 will not make any break changes on ERC721. It can be merged with ERC721, or deployed as an independent contract and associated with an existing ERC721 contract.

Yes, AI721 is essentially NFT(non-fungible token) with AI capabilities.

We call it AI-NFT for short.

Why we need AI721

LLMs enable everyone to create AI personalities with any prompts. You can create any kinds of characters with different personalities with some tools like However, it's centralized, non-transparent, and unable to achieve consensus.

If we want to describe AI personalities in a decentralized environment , and access it without relying on any centralized 3rd parties, we need to define a standard to implement this process, ensuring that anyone can obtain the same AI personality through the same protocol.

AI721 is the key to make it happen.

How it works

In the AI721 protocol, the AI Personality of an AI-NFT is defined by these core components below:

  • LLM: The large language model that can comprehend and generate human language text. It provides basic AI capabilities for your AI-NFT.

  • Trait model: A collection-level function that inputs the AI-NFT's traits and outputs a set of prompts, defining the independent personality that is only related to its traits, as an initial state.

  • Memory: The memories, or knowledge that form the final AI personality, manifested through promps in conversation with the AI agent.

    • Public Memory: A predefined and public memory that anyone can use to form an AI personality.

    • Private Memory: A user-level private memory that shared with your AI agent, can be only read by the NFT owner.

  • Workflow: A workflow is defined as a structured sequence of operations that organizes tasks into manageable steps. Compatible with the data structure of the workflow on

How to describe AI Agent on chain?

AI Metadata Standard

AI721 defines a new kind of JSON metadata to describe an AI personality. The metadata standard includes three important parameters to index the AI personality. This metadata is different from the one for ERC721, and can be stored on-chain or off-chain.




The uri to access LLM that is stored in off-chain storages.

Check URI Format Supported.



The prompt template that generates the final prompts from the traits and other storyline of AI-NFT, by using placeholders of the form ${trait_type} , to define an independent and unique personality. Check Trait Model Example.



(Optional) The uri to access the list of public memories needed to form an AI agent.

The application layer can choose from these memories to create specific AI agents, like a virtual lawyer, math teacher, or virtual girlfriend/boyfriend. Check URI Format Supportedand Memory Structure.



(Optional) The uri to access the workflow file that describes the action chain of an AI agent.

Trait Model Example

A typical traits array from the metadata of a certain NFT is like:

  "attributes": [ 
      "trait_type": "Base", 
      "value": "Starfish"
      "trait_type": "Eyes", 
      "value": "Big"
      "trait_type": "Mouth", 
      "value": "Surprised"

We need to convert it to:


Then the trait_model could be like:

"You have these traits below:
- Base: {{Base}}, means the basic nation
- Eyes: {{Eyes}}, means what your eyes are look like
- Mouth: {{Mouth}}, means what your month are look like
- Influence: High, means your reputation
- Personality: Analytical, Pragmatic, Influential

We use Jinja2 as the templating engine for trait model.{{Base}}, {{Eyes}}, {{Mouth}} are all placeholders in the template that will be replaced with traits.Base, traits.Eyes and traits.Mouth of a specific AI-NFT.

On the application side, after you retrieve trait_model , you can render the final prompts by filling in specific traits data with Jinja2.

Memory Structure

A single memory is described as a JSON file.




The name of this memory



(Optional) The description of this memory



(Optional) The logo url or the memory



The serialized form of this memory.

Workflow Structure

A workflow is a configuration described as a JSON file.



The engine that can run the workflow given by For example, If the workflow originates from, or is compatible with this platform, set this field to “dify”.



(Optional) The engine version.



The DSL to describe the workflow of the AI agent.

Based on this design, AI721 can support various current and future AI agent workflow protocols, and use the corresponding engines to run AI agent services.

URI Format Supported



Access a given LLM by calling APIs hosted by a party. This is only applicable for llm.

For examples:

  • "gpt-4" means gpt-4 model or its extended versions provided by OpenAI.

  • "gemini-1.0" means gemini-1.0 model or its extended versions provided by Google.

  • "any" means any LLMs can be used.

*The name must follow the naming conventions used in the API parameters provided by major model manufacturers.

The application layer can determine how to access these models' API endpoints, for example, by registering a developer account at


The resources stored in IPFS/Filecoin. e.g. ipfs://QmbWqxBEKC3P8tqsKc98xmWNzrzDtRLMiMPL8wBuTGsMnR


The resources stored in Arweave. e.g. ar://k2g_3kYsPx-meD-AlyHhkUDYbUczlZ-M-bsKO6_oqY4


An typical http(s) url to locate resources.

... More formats from DePIN partners

How to design a workflow

Check the section.

How to access the AI agents described by AI721?

  1. Call tokenURI by tokenId of an AI721 contract, and pull the AI metadata of the NFT.

  2. Retrieve the LLM by llm and run as a service;

  3. Retrieve the trait model by trait_model. Input the traits set of the NFT to the model to obtain a set of prompts, and then integrate these prompts into the LLM you get in the previous step;

  4. (Optional) Retrieve one of the public memories from the public_memories , and integrate it into the LLM you get in the previous step. This allows you to quickly create a predefined AI personality.

  5. (Optional) Retrieve workflow and load it locally for the application layer to trigger at any time. If the workflow involves external LLMs, you can either replace them with the dedicated LLM from Step 4th or continue using the external LLM interface described by the DSL.

After these steps, you'll get your AI agent ready to use, with a pre-defined personality.

Implementing token URI of AI metadata

As what you have done for ERC721, you MUST implement tokenURI and return a URI where we can find the metadata.

Best Practice

You can deploy an AI721 contract inherit from ERC721 to issue AI-NFT. In this way, you MUST directly add new AIMetadata fields to ERC721 metadata. The application layer will call tokenURI(uint256 _tokenId) to access the metadata of an NFT and its AI personality directly.

Merged Metadata Example

// Merge AIMetadata with ERC721 metadata 
  // ERC721 metadata fields
  "description": "Friendly OpenSea Creature that enjoys long swims in the ocean.", 
  "external_url": "", 
  "image": "", 
  "name": "Dave Starbelly",
  "attributes": [ 
      "trait_type": "Base", 
      "value": "Starfish"
      "trait_type": "Eyes", 
      "value": "Big"
      "trait_type": "Mouth", 
      "value": "Surprised"
  // AI fields
  - Base: {{Base}}
  - Eyes: {{Eyes}}
  - Mouth: {{Mouth}}
  - Influence: High
  - Personality: Analytical, Pragmatic, Influential

We recommend the application layer to firstly access AI metadata embedded in ERC721 contract to query AI personalities. If not found, then use AILinker to query.


Last updated