I’m a substrate enthusiast and like to vibe code blockchain and MCP (model context protocol, a tool that can be use by an AI).
I am looking to build a system that can be used as a public key infrastructure for building deliverables through AI assistance. I already build the the substrate blockchain and most of the dapp that I put in a MCP, so the UI of the dapp is a natural conversation through the AI. It could help web3 adoption by traditional business.
Once I finished my stack, I want to see if a decentralized network of local and specialized AI agent(like Llama or mistral) can be an alternative to huge centralized model.
So I’m looking for fellow AI and blockchain enthusiast with local AI model that want to experiment.
If anyone is interesting, I can show freely how to vibe code cleanly a substrate blockchain with no prior rust or frame knowledge in a few hours. We can build together a governance pallet exploiting the democratic system of Athenian boulé for example. It’s all about prompting the AI for giving it a bias in pattern discovery then chiselling the algorithm into the code.
Be sure share all the projects vibecoding their blockchain so I can avoid them.
I use AI tool daily and I will never trust blindly them on any security related context. If you are already a decentralized protocol expert, it is a tremendous productivity booster, otherwise it is a recipe for disaster.
If you can create permissions/rules with constraints on what the AI can or cannot do and start with basic parachain template generation from audited parachain templates.
Then offer basic configuration: weights, transaction fees, native token configuration, adding common signed extensions, adding pre-approved audited pallets, basic XCM configurations, creating patches for release upgrades, and adding governance using only existing audited pallets. And of course all changes would be made in the form of a PR that would need to be code reviewed.
What would be cool would be to have the AI “eat its own dog food” — have it generate a template then deploy it with its own keys. If the AI model is good it can then maintain the parachain by applying releases in the form of runtime upgrades. If the AI fails the parachain will get bricked and it will have to start over again from scratch. Make it like a game. Eventually after so many iterations it will succeed. Reminds me of when President John F. Kennedy challenged the United States in 1961 to land a man on the moon and return him safely. It would be a good marketing stint for Polkadot — sending DOT to the moon. Also reminds of the SpaceX rockets that sometimes and millions of dollars are lost but the rocket made a bit more progress than its previous iteration.
A better approach may be to create runtime level MCP Make Polkadot AI-friendly by creating a communication layer for agents. Remember an MCP can go both ways (read and write) so agents can read from a node and perform actions on Polkadot like calling a runtime level API.
Sure, starting vibe coding after doing the zero to hero tutorial will result in a monstrosity with an optimizedHybridRNG of 6 local sources of imaginary randomness and an OCWFactory at start.
However it is too an amazing teaching tool for those who wanna learn, but it is slow. It is not a magic tool but a high level programming syntax.
If the algorithm is clear in your head , if the code is auto explicit, if you can detect early hallucination pattern and if you know the basis of what good code is, then it’s awesome, otherwise it will be at best an elegant solution to a problem that doesn’t exist.
Yes, I use Stdio for the MCP that contain the DApp, it allow both side communication. There will be a cool feature that will be release in the coming month for MCP that is elicitation.
With elicitation, when something happen in the server/DApp inside the MCP, an elicitation is send to the AI, consisting of instruction and a Json Schema. Then the AI will execute the instruction, create an object that is validated by the schema and send it to the DApp.
I don’t think you can trust an AI to create blockchain unsupervised, there are not that efficient and prone to hallucination. Crypto algorithm are out of range however they are good to see graphs.
and millions of dollars are lost but the rocket made a bit more progress than its previous iteration.
Make Polkadot AI-friendly by creating a communication layer for agents. Remember an MCP can go both ways (read and write) so agents can read from a node and perform actions on Polkadot like calling a runtime level API.