Introducing Flow AI Tools: The Claude Code Plugin for Web3 Development
The promise of AI-assisted programming is simple: you write the prompts, the model ships the code. The reality for Web3 is often something more along the lines of: the model writes something that looks right, you try to deploy it to testnet, and it fails. The LLM tells you the app is done, but the model never actually ran what it wrote. Or it generates 300 lines to recreate something that already exists in two. These experiences are not only frustrating, they waste time and money.
The security cost of vibe coding lies in not having full knowledge of what's being produced. The model is fluent, fast, and confidently wrong about insecure code patterns or outdated versions it's been trained on. Every minute spent debugging a hallucinated entitlement is a minute not spent on what actually matters: shipping a product and getting it in front of users. Distribution is where decentralized apps win or lose. Babysitting the model is dead weight.
The Flow Claude Code Plugin, published by Flow and supported through community contributors, gives Claude the context it's missing to build useful decentralized applications. That includes the bleeding-edge language spec for Cadence, the current standards used across the ecosystem, and the patterns that actually compile.
The plugin makes it possible to build decentralized apps fast that are largely impossible elsewhere. Digital financial assets, game characters, art, tickets, and deeds become items that behave like physical possessions, with advanced metadata that can grow and expand as the application evolves. Face ID to sign transactions instead of a seed phrase. Modding and extending any existing on-chain app or game with prompts. AI agents and DeFi apps with built-in safeguards, where a wallet can be scoped to "move up to $100 a day, nothing else, revocable instantly," or a dollar-cost-averaging bot can buy $50 of crypto every morning and revert the whole transaction if it ever overpays, without relying on anything offchain.
How to install the Flow Claude Code Plugin
With a one-line install and plain-English prompts to set it up, this plugin makes it easier than ever to build and deploy apps on Flow. Anyone can build a marketplace that operates 24/7, launch on-chain tickets, or create novel options for DeFi yield. All on the same network that Disney, the NBA, the NFL, and Ticketmaster trust.
Install by running this one line in your terminal:
sh -ci "$(curl -fsSL https://raw.githubusercontent.com/onflow/flow-ai-tools/main/scripts/install.sh)"11 skills covering every part of your decentralized app
| Skill | What it does |
|---|---|
| cadence-lang | The current Cadence language spec, complete and up to date. |
| cadence-tokens | NFT and FT contract standards, plus metadata patterns so minted assets show up richly across marketplaces with images, traits, and onchain royalties. |
| cadence-scaffold | Generates contracts and transactions with secure-by-default patterns: entitlements, capability scoping, and post-conditions. Sensitive functions are locked down from the first commit. |
| cadence-audit | Reviews code across security, correctness, quality, and computation cost, returning severity-rated findings. |
| cadence-testing | The Cadence test framework, flow test, and the built-in Test contract API. Helps Claude actually test its own work before claiming it's done. |
| flow-cli | Full Flow CLI reference: account management, querying chain state, deployment workflow, script recipes, and Cadence MCP server setup. |
| flow-project-setup | flow.json structure, FCL frontend integration, deployment across emulator, testnet, and mainnet. |
| flow-dev-setup | Day-zero machine setup. Flow CLI, emulator with fork mode, VS Code Cadence extension, dev wallet, testing framework, and EVM tooling (Hardhat, Foundry, Remix). |
| flow-react-sdk | Hooks for auth, contract reads, events, and transactions, plus cross-VM hooks for Flow EVM. Wires up the whole app across Cadence and Solidity. |
| flow-defi | DeFi protocol architecture: COA patterns, MEV-free EVM, cross-VM atomicity, lending math, AMM types, liquidity bootstrapping, and the Flow DeFi ecosystem map. |
| flow-tokenomics | Token economics and launch strategy: first principles, proven patterns, failure case studies, a 12-week TGE playbook, and regulatory considerations. Note: AI and skills do not give financial advice. |
Higher security when vibe coding
Having both DeFi skills and security auditing skills available in one plugin minimizes the chance of creating projects with dangerous, exploitable code. Ask Claude to scaffold a lending vault and cadence-scaffold writes the contract, cadence-lang enforces the entitlements, cadence-audit reviews what came out, and flow-defi checks the architecture against patterns that have actually shipped, including health factor curves, kink models, and MEV-free execution on Flow EVM.
You can never fully ensure with skills alone that there is no possible vulnerability. Even with expensive human audits there is always risk. But using AI to pre-vet production code is one of the most cost-effective and time-efficient ways to increase the hygiene of your code.
It won't replace deploying to testnet before mainnet, and it won't replace a final human review when real value is on the line. Treat this like pair-programming with a senior developer who's shipped a lot of Flow code, not autopilot.
Optimized skills that self-trigger at the right time
Each skill's frontmatter declares when Claude should pull it in, so you don't manually invoke anything. Ask "review this NFT contract" and cadence-audit and cadence-tokens activate; the React SDK guide stays dormant. Skills also list their companions, so when you're scaffolding a DeFi vault, the model consults cadence-lang for syntax, cadence-audit for review, and flow-defi for architectural choices. This happens without you orchestrating it or wasting tokens filling up limited context space.
The skills inform Claude what good Cadence looks like. The Cadence MCP server, also installed by the one-liner, lets Claude actually do things on chain: type-check what it just wrote, fetch the source of any deployed contract on Flow, run a security review pass, and execute read-only scripts against testnet or mainnet. Ask "what's the current FlowToken supply?" and the model can run a script against mainnet instead of guessing. Ask "is this contract using the same Vault pattern as the new DeFi product that just launched?" and the model can fetch that project's contract addresses and source code, then compare. The result is enhanced verification before shipping, better chain reads, and fewer hallucinations from outdated training data.
Contribute to the best place to build with AI
Have an idea for another killer skill you want to contribute to the developer community? Discovered something the audit skill missed, or noticed a deprecated pattern? Open an issue or PR at github.com/onflow/flow-ai-tools and help make this the best Web3 development plugin for Claude Code.
For all Flow developer documentation, visit (or send your agents to) developers.flow.com.
