AI Is Reshaping Crypto Development: At Least 14% of Top Projects Now Code with AI

AI-assisted coding is no longer a side topic in crypto development. It is beginning to show up inside the daily engineering workflows of major protocols, wallets, security tools, and documentation projects.

I analyzed 1,000 of the most-starred crypto-related repositories on GitHub, using repository data provided by Electric Capital’s Crypto Ecosystems dataset. Among those repositories, at least 137 projects showed explicit, verifiable evidence of AI-assisted coding.

That means roughly 14% of the top crypto repositories are already using AI coding tools in ways that leave public traces.

This should be read as a conservative lower bound, not a full measurement of adoption. The research only counted evidence visible in commits, pull requests, or other repository activity. Local IDE usage, private agent workflows, and unmentioned AI assistance would not be captured.

What AI Is Being Used For

AI tools are not only being used for simple code completion. Across the repositories that showed evidence of AI usage, the recurring tasks included:

• Automated code review.
• Bug and vulnerability detection.
• Coding assistance and completion.
• Translation of multilingual components.
• Documentation and specification maintenance.

The pattern is clear: AI is entering both product code and the surrounding engineering system. It is not limited to scripts at the edge of a repository.

Among projects with visible AI usage, Claude and GitHub Copilot dominate. Together, they account for roughly 87% of observed AI tool usage. Other tools, including OpenAI’s Codex, Cursor, and Google Gemini, also appear, but they currently represent a smaller share in this dataset.

Where AI Shows Up

AI usage is already visible across several layers of the crypto stack.

Layer 1 and Layer 2 protocol repositories showed signs of adoption, including:

• ethereum/go-ethereum
• aptos-labs/aptos-core
• MystenLabs/sui
• near/nearcore
• stellar/stellar-core
• ethereum-optimism/optimism

Wallet projects also showed clear evidence:

• MetaMask/metamask-extension
• rainbow-me/rainbow
• WalletConnect/walletconnect-monorepo
• BlueWallet/BlueWallet

Security and auditing repositories appeared in the dataset as well:

• trailofbits/algo
• crytic/solc-select

Documentation and specification repositories were another notable category:

• ethereum/EIPs
• ethereum/ethereum-org-website
• MetaMask/metamask-docs
• duneanalytics/spellbook
• ethereum/execution-specs
• foundry-rs/book

This matters because documentation, specifications, and ecosystem knowledge bases are not peripheral in crypto. They are part of the coordination layer of open-source networks. If AI is shaping how these artifacts are maintained, it is also shaping how the ecosystem explains itself.

Organizations Are Mixing Tools

One useful signal is that organizations are not simply adopting one AI tool everywhere. They often match tools to a repository’s stack, workflow, or task type.

For example, MetaMask showed different tool usage across related repositories:

• metamask-extension used Cursor.
• metamask-docs used Cursor.
• metamask-mobile used Claude.

a16z-related repositories showed a similar pattern:

• helios used Claude.
• halmos used Copilot.
• jolt used OpenAI.

Coinbase-related projects also showed tool diversity:

• x402 used Cursor.
• onchainkit used Copilot.

This suggests a more pragmatic phase of adoption. Teams are not treating AI coding as a single product category. They are assembling a toolchain.

Methodology

The research followed a simple, conservative process.

First, I collected crypto-related GitHub repository data from more than half a million records in Electric Capital’s dataset. I then gathered star counts and selected the top 1,000 non-archived repositories by popularity.

Second, I analyzed the latest 200 commits and the latest 200 merged pull requests on each repository’s main branch. The search focused on keywords associated with AI coding agents and tools, including Claude, Copilot, Codex, Cursor, and Gemini.

Third, I manually checked the initial matches to reduce false positives. For example, “Gemini” can refer to Google’s AI model, but it can also refer to the crypto exchange.

Even with manual verification, this kind of research cannot claim absolute accuracy. It is better understood as an open, reproducible snapshot of visible AI usage in crypto repositories.

Why This Matters

Crypto is a developer-driven industry. Protocol quality, security response, documentation accuracy, and ecosystem velocity all depend on how fast and how well teams can ship software.

If AI tools improve review speed, reduce common bugs, accelerate documentation, or help teams respond to new vulnerabilities, they can change the competitive dynamics of crypto development.

The more important point is not that AI will replace crypto developers. It is that AI-assisted teams may develop a compounding operational advantage. They can move faster, maintain more surface area, and coordinate larger codebases with less friction.

For projects and investors, AI coding adoption may soon become a signal worth watching. Not because every repository needs to mention AI in its README, but because the engineering workflows behind successful crypto projects are changing.

The visible 14% figure is only the beginning. The hidden number is almost certainly higher.

Originally published as an open research note: state_of_ai_coding_in_crypto_en.md.

All detailed data is available at: crypto-ai-coding-report.