Solana AI Kit
GitHub: solanabr/solana-ai-kit Status: Active development, public release Maintained by: @kauenet
Why It Exists
AI coding assistants have a Solana problem. Their training data is months or years behind a fast-moving ecosystem, which means they produce code that looks right but isn't: unchecked arithmetic that silently overflows, missing account owner validation, deprecated web3.js 1.x patterns when the project uses @solana/kit, hallucinated API signatures for Anchor versions that no longer exist. The AI confidently writes the wrong thing, and the developer has to know enough to catch it.
The second problem is context waste. A general-purpose assistant loads everything and filters nothing. Asking about a Pinocchio CPI pattern pulls in irrelevant React knowledge. Asking about token metadata loads game development rules. Every token spent on context that does not apply to the current task is a token that cannot go toward solving the actual problem.
Solana AI Kit (formerly solana-claude-config) solves both. It is an AI-coding configuration — Claude Code and Codex compatible — that encodes deep Solana domain knowledge (current APIs, correct security patterns, language-specific rules) and delivers that knowledge efficiently through a token-aware architecture. The right context loads for the right task, and nothing else does.
Design Philosophy
The configuration is built around a single principle: load only what the current work requires.
Short CLAUDE.md (~110 lines) as user-message context. Only essential rules, security principles, and workflow instructions live here. The file ships as a token budget, not a reference manual.
Progressive skill loading via .claude/skills/. Specialized knowledge — DeFi protocol patterns, game development conventions, security audit checklists — loads on demand through slash commands or agent invocations, not on every prompt.
Lazy rule loading via .claude/rules/. Language-specific standards load only when working in that language. The Rust rules are invisible during a TypeScript refactor. The C# rules are invisible during an Anchor program audit.
Agent-scoped context. Each of the 15 agents carries only the tools and knowledge relevant to its role. The unity-engineer has C#/.NET rules and Solana.Unity-SDK patterns. The pinocchio-engineer has zero-copy patterns and manual validation. Neither carries the other's context.
CLAUDE.local.md for scratch notes. Private, gitignored, written by Claude during sessions. Project observations and session summaries that should not be shared with the team.
Monorepo support. Subdirectory CLAUDE.md files for scoped architecture decisions. Claude Code loads them automatically when working in that directory.
What You Get
15 Specialized Agents
solana-architect
System design, account structures, PDA schemes, cross-program composability
anchor-engineer
Anchor program development with IDL generation and standardized patterns
pinocchio-engineer
CU optimization with zero-copy framework (80–95% CU reduction vs Anchor)
defi-engineer
DeFi protocol integration — Jupiter, Drift, Kamino, Raydium, Orca, Meteora
solana-frontend-engineer
React/Next.js dApp frontends with wallet adapter integration
game-architect
On-chain game design, Unity architecture, PlaySolana ecosystem
unity-engineer
Unity/C# with Solana.Unity-SDK, wallet integration, NFT display
mobile-engineer
React Native and Expo for Solana mobile dApps
rust-backend-engineer
Async Rust services with Axum/Tokio for Solana backends
devops-engineer
CI/CD, Docker, monitoring, RPC management, Cloudflare Workers
token-engineer
Token-2022 extensions, token economics, transfer hooks, compliance
solana-researcher
Deep research on protocols, SDKs, and ecosystem tools
solana-qa-engineer
Testing (Mollusk, LiteSVM, Surfpool, Trident), CU profiling, fuzzing
tech-docs-writer
READMEs, API docs, integration guides, architecture documentation
solana-guide
Developer education, tutorials, learning paths
30 Workflow Commands
Building
/build-program— Build and verify Solana programs/scaffold— Project scaffolding from templates/build-app— Full-stack application setup/build-unity— Unity game project scaffolding
Testing and Quality
/test-rust— Rust test runner with coverage/test-ts— TypeScript test runner/test-dotnet— .NET/Unity test runner/test-and-fix— Run tests and automatically fix failures/audit-solana— Security audit for on-chain code/profile-cu— Compute unit profiling and optimization/benchmark— Performance benchmarking/diff-review— Code review and AI slop detection/audit-infra— Infrastructure-first security audit: secrets, supply chain, CI/CD, LLM/skill security, OWASP, STRIDE/product-review— Product quality review with an 8-dimension scorecard/debug-user-tx— Replay a user's failing transaction against forked cluster state and map the error to source
Deployment
/deploy— Devnet and mainnet deployment with confirmation gates/setup-ci-cd— CI/CD pipeline configuration/setup-mcp— MCP server configuration and key management
Workflow
/quick-commit— Branch creation and commit automation/explain-code— Explain unfamiliar Solana patterns/write-docs— Generate documentation from source/plan-feature— Architecture planning before implementation/generate-idl-client— TypeScript client generation from IDL/migrate-web3— Migrate web3.js 1.x code to @solana/kit/update— Update configuration and skills to latest/resync— Resync agent and rule definitions/cleanup— Remove stale context and temporary files/commit-claude-config— Version the kit configuration in git/doctor— Read-only health check for the dev environment and config, with one fix-it command per failure/dream— Memory consolidation: dedupe, prune, and re-rank session learnings
18 External Skill Submodules
Each submodule is a git submodule sourced from an authoritative provider, keeping domain knowledge current without embedding it directly into the configuration.
solana-dev
Solana Foundation
Official Solana development patterns — programs, frontend, testing, security
sendai
SendAI
DeFi protocol integrations — Jupiter, Raydium, Kamino, perps, cross-chain, oracles
solana-game
Superteam Brazil
Unity game development for Solana — PlaySolana, PSG1
cloudflare
Cloudflare
Edge infrastructure — Workers, Agents SDK, MCP servers
trailofbits
Trail of Bits
Security auditing and vulnerability scanning
qedgen
QEDGen
Formal verification with Lean 4 theorem proving
solana-mobile
Solana Mobile
Mobile Wallet Adapter, Genesis Token, SKR address resolution
colosseum
Colosseum
Startup research, idea validation, hackathon projects
safe-solana-builder
Community
Security-first code generation with 70+ audit-derived rules
vercel
Vercel
Vercel deployment, Next.js, AI SDK, v0, edge functions
solana-new
SendAI
Idea-to-launch journey skills plus idea datasets and knowledge base
ghostsecurity
Ghost Security
AppSec skills — SAST criteria, SCA, secrets, validation
defending-code
Anthropic
Vulnerability-discovery reference harness and defensive-security skills
jupiter
Jupiter
Official Jupiter skills — Ultra swap, Lend, swap migration
metaplex
Metaplex
Official Metaplex skills — Core, Token Metadata, Bubblegum, Candy Machine
helius
Helius
Official Helius infrastructure skill plus SVM internals
quicknode-anchor
QuickNode
Anchor and financial-math reference files (quarantined — references only)
eth-to-sol
Solana Foundation
EVM/Solidity to Anchor two-pass porting
7 MCP Server Integrations
Helius
60+ tools for RPC, DAS API, webhooks, priority fees, and token metadata — real-time on-chain data without leaving the editor
solana-dev
Solana Foundation official MCP with current docs, guides, and API references — the AI never hallucinates deprecated APIs
Context7
Fetches current documentation for any dependency, not training-data snapshots
Playwright
Browser automation for dApp testing — opens your frontend, connects wallets, and verifies flows in a real browser
context-mode
Compresses large RPC responses and build logs — saves context window for actual work
memsearch
Persistent memory with semantic search — remembers project context across sessions
Surfpool
Agent-driven local validator and mainnet-fork control — spin up forked cluster state for integration testing without leaving the editor
Language-Specific Rules
Enforced coding standards that load only when relevant to the current task:
Rust — Checked arithmetic, proper error propagation, no
unwrap()in production codeAnchor — Account validation constraints, PDA bump storage, CPI target validation, account reloading after CPI
Pinocchio — Zero-copy access patterns, manual
TryFromvalidation, single-byte discriminatorsTypeScript — Type safety (
no any), async/await patterns, wallet adapter integration, BigInt for u64C#/.NET — Unity MonoBehaviour conventions, .NET 9 standards, Solana.Unity-SDK integration patterns
Agent Team Patterns
Multi-agent workflows that coordinate specialized agents through a complete development cycle. Create a team with natural language: "Create an agent team: solana-architect for design, anchor-engineer for implementation, solana-qa-engineer for testing".
program-ship
architect → engineer → QA → deploy
Ship a complete Solana program
full-stack
architect → engineer → frontend → QA
End-to-end dApp development
audit-and-fix
QA → engineer → QA
Security audit with automated fixes
game-ship
game-architect → unity-engineer → QA
Unity game with on-chain state
research-and-build
researcher → architect → engineer
Research-driven implementation
defi-compose
researcher → defi-engineer → QA
Multi-protocol DeFi integration
token-launch
token-engineer → QA → deploy
Token creation with extensions
Installation
Fork the template — the recommended approach for new projects. Fork solanabr/solana-ai-kit on GitHub, customize the CLAUDE.md for your project, and initialize the skill submodules.
One-liner install — for adding the configuration to an existing project:
If the aikit.superteam.codes domain is not reachable, the documented fallback is:
Claude Code plugin — installs the core kit (agents, commands, MCP servers) through the plugin marketplace:
The plugin carries the core kit only; the full install.sh route also brings the lazy-loaded language rules, the curated permissions allowlist and sandbox policy, and the 18 ext/ skill submodules.
Manual setup — clone the repository and copy the .claude/ directory and CLAUDE.md to your project root.
Agent export — the --agents flag exports agent definitions for use with non-Claude tools that support the agent specification format.
Modern Stack (2026)
Programs
Anchor 1.0+ / Pinocchio / Rust 1.82+
Frontend
Next.js 15 / React 19 / @solana/kit
Testing
Mollusk / LiteSVM / Surfpool / Trident
Mobile
React Native / Expo / Solana Mobile SDK
Games
Unity 6+ / Solana.Unity-SDK / PlaySolana
Backend
Rust (Axum 0.8+ / Tokio 1.40+ / sqlx) / Helius API
Edge
Cloudflare Workers
Credits
This project would not be possible without the organizations that publish and maintain the skill submodules it depends on. Thanks to the Solana Foundation for official development patterns and the solana-dev MCP server, SendAI for the AI agent framework, Trail of Bits for the security research this project draws from, Cloudflare for edge deployment patterns, Colosseum for hackathon tooling, QEDGen for formal verification work, Solana Mobile for mobile SDK patterns, the safe-solana-builder community for safe coding conventions, Vercel for deployment patterns, Jupiter, Metaplex, and Helius for their official ecosystem skills, Ghost Security for AppSec skills, Anthropic for the defending-code reference harness, QuickNode for Anchor reference material, and Superteam Brazil for the game development integration that started this project.
Last updated
Was this helpful?