Solana AI Kit

Production-ready Claude Code configuration for full-stack Solana development. Combines best practices from multiple sources into an agent-optimized, token-efficient config you can install and adapt to your specific project.

The idea here is to provide a generic CLAUDE.md that relies on subagents to plan and execute actions, dynamically loading markdown files, saving tokens and context in the end of the day. This config fully leverages the official Claude Code config recommendations:

• Rules are only loaded whenever specific file types are involved;
• SKILL.md is a mega hub to dynamically-disclosed skill files that are directly fetched from the best skill repos distributed across the ecosystem (Solana Foundation, Colosseum, Solana Mobile, SendAI, etc);
• Plus, its CLAUDE-solana.md is than half the size of the usual CLAUDE.md, leaving space for its self improvements programmed into the agents, noting and learning from anti-patterns, errors, recurrency and more. For less important notes, CLAUDE.local.md is constantly maintained by agents as well and, on monorepos, per-folder CLAUDE.md is also maintained.

Current multi-agent workflow favors monorepos, so we use a single CLAUDE.md/config for the whole project while leveraging agents and context-specific skills to solve each step of builder flow.

If you installed manually, remember to rename ./CLAUDE-solana.md back to ./CLAUDE.md, as the current top-level CLAUDE.md file is focused on maintaining the repo itself.

What This Is

A complete .claude/ configuration that turns Claude into a Solana development expert with:

• 15 specialized agents for different tasks (architecture, Anchor, Pinocchio, DeFi, tokens, frontend, mobile, backend, DevOps, QA, docs, games, Unity, learning, research)
• 29 workflow commands for building, testing, deploying, profiling, migrating, and committing
• 7 MCP server integrations for on-chain data (Helius), Solana docs (solana-dev), library docs (Context7), browser automation (Playwright), context optimization (context-mode), persistent memory (memsearch), and local-validator / mainnet-fork control (Surfpool)
• Agent teams for multi-step workflows (architect → engineer → QA)
• Progressive skill loading that only loads context when needed (saves tokens)
• Auto-loading rules that enforce best practices based on file patterns

Quick Start

# Option 0: Fork Template (Github UI)
claude -m "/cleanup"  # then start claude code running /cleanup so top-level directory of your project isn't polluted


# Option 1: One-liner installer (Claude Code)
curl -fsSL https://raw.githubusercontent.com/solanabr/solana-ai-kit/main/install.sh | bash

# Option 2: One-liner installer (Codex, Opencode, everything else)
curl -fsSL https://raw.githubusercontent.com/solanabr/solana-ai-kit/main/install.sh | bash -s -- --agents

# Option 3: Manual setup
git clone --recurse-submodules https://github.com/solanabr/solana-ai-kit.git
cp -r solana-ai-kit/.claude /path/to/your-project/
cp solana-ai-kit/CLAUDE-solana.md /path/to/your-project/CLAUDE.md
cd /path/to/your-project && git submodule update --init --recursive

# Start Claude Code
claude

--agents Flag

Pass --agents to install everything into .agents/ instead of .claude/. Same content, different directory name. Useful when .claude/ is already taken or for non-Claude AI tools (Cursor, Windsurf, Copilot, etc.).

bash install.sh --agents /path/to/your-project

To update an --agents install:

# Run from your project root:
bash .agents/bin/update.sh

MCP Setup (Optional)

After installation, configure MCP servers for enhanced capabilities:

# In your project with Claude Code running:
/setup-mcp

This guides you through API key configuration for Helius, Context7, and other MCP servers.

Install as a Claude Code plugin

solana-ai-kit is also its own Claude Code marketplace serving one core plugin. From inside Claude Code:

/plugin marketplace add solanabr/solana-ai-kit
/plugin install solana-ai-kit@stbr

The plugin ships the core kit: the 15 agents, 29 commands, the local go-to-market + registry skills (idea-sprint, pitch-deck, hackathon, registry), the 7 MCP servers, and the dev hooks (banner, formatter, pre-deploy/commit gates). Commands and skills are namespaced — /deploy becomes /solana-ai-kit:deploy.

What the plugin cannot carry (Claude Code plugins are plain git clones — they can't init submodules or ship a permissions/sandbox policy), so these stay exclusive to the full install (install.sh):

• the lazy-loaded .claude/rules/* code-style law (Rust, Anchor, Pinocchio, TypeScript, .NET)
• the curated permissions allowlist + sandbox policy
• the 18 ext/ skill submodules (protocol, security, infra, ecosystem depth)

For protocol-skill depth in plugin form, add the upstream marketplaces instead (routing, not copying) — e.g. /plugin marketplace add sendaifun/skills, ghostsecurity/skills, trailofbits/skills, cloudflare/skills. The plugin's skill hub and the registry skill list the current targets.

The two paths are complementary: individuals wanting Solana agents/commands across all their projects → plugin at user scope; project teams working in one repo → install.sh (full kit). If you enable both in the same project, /doctor warns about duplicate commands/hooks/MCP and advises picking one. Plugin auto-update is off by default for third-party marketplaces — refresh with /plugin marketplace update + claude plugin update.

Key Features

External Skill Submodules

Agent Teams

Each agent loads its own specialized context on invocation:

"Use solana-architect to design the vault program"
"Use anchor-engineer to implement the deposit instruction"
"Use defi-engineer to integrate Jupiter swaps"
"Use solana-qa-engineer to write comprehensive tests"

Claude will spawn each specialized agent by itself based on context and, with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS enabled, orchestrate multi-agent workflows:

"Create an agent team: solana-architect for design, anchor-engineer for implementation, solana-qa-engineer for testing"

Recommended team patterns:

MCP Server Integrations

Pre-configured MCP servers in .mcp.json (API keys go in .env):

Token-Efficient Design

• CLAUDE.md is delivered as a user message (not system prompt) — shorter = better adherence
• Skills load progressively (not all at once)
• Agents reference skills instead of duplicating content
• Path-scoped rules lazy-load on file read (zero startup cost)
• CLAUDE.local.md for private scratch notes (gitignored, never shared)
• Subdirectory CLAUDE.md files lazy-load in monorepos
• Decision frameworks live in agents, not global context

Modern Stack (2026)

Extended Skills & Add-on Registry

Beyond the bundled submodules above, the kit ships a curated catalog of opt-in skills, plugins, and MCPs that are not installed by default — the agent installs them only on your request, at your own expense. The set is deduped against solana-new's vendored ecosystem catalogs so it surfaces net-new, high-signal picks rather than restating what's already reachable.

Featured add-ons by domain:

• Claude-official: anthropics/skills · anthropics/claude-code plugins (non-OSI license; overlaps /diff-review + cso)
• Dev-workflow: wshobson/agents · gsd-build/get-shit-done
• Frontend/Design: zarazhangrui/frontend-slides · uxKero/anydesign · dylantarre/animation-principles
• UX/Writing: content-designer/ux-writing-skill · cuellarfr/design-skills
• Testing: SawyerHood/dev-browser · conorluddy/ios-simulator-skill
• Data: K-Dense-AI/scientific-agent-skills · Nansen (paid analytics MCP)

Where we scout new tools (aggregators, not installable): ComposioHQ/awesome-claude-skills · travisvn/awesome-claude-skills · davepoon/buildwithclaude · hesreallyhim/awesome-claude-code · VoltAgent/awesome-claude-code-subagents.

For broader Solana coverage, see solana-new's vendored catalogs at ext/solana-new/cli/data/ (MCPs, skills, clonable repos).

See skill-registry.json for the complete extended catalog — every entry with its install command, license, and safety caveats.

Repository Structure

.
├── CLAUDE.md                    # Main hub - Claude reads this first
├── README.md                    # This file
├── .mcp.json                    # MCP server configurations (project root)
├── install.sh                   # One-liner installer
├── update.sh                    # Deprecation wrapper → .claude/bin/update.sh
├── validate.sh                  # Config integrity checker
├── LICENSE                      # MIT
├── tests/                       # Config integrity test suite
├── .github/workflows/
│   ├── ci.yml                       # PR validation
│   ├── claude-code.yml              # Claude Code action template
│   └── claude.yml                   # @claude mention responder (issues/PRs)
└── .claude/
    ├── VERSION                  # Semver version (e.g. 1.4.0)
    ├── agents/                  # 15 specialized agents
    ├── bin/
    │   ├── update.sh                # In-place update from upstream
    │   └── resync.sh                # Submodule resync script
    ├── commands/                # 29 workflow commands
    ├── skills/                  # Progressive-loading knowledge
    │   ├── SKILL.md                 # Unified hub routing to all skills
    │   ├── ext/                     # External skill submodules
    │   │   ├── solana-dev/              # Solana Foundation dev skill (core)
    │   │   ├── sendai/                  # SendAI protocol skills (DeFi)
    │   │   ├── solana-game/             # Solana game skill (Unity, PSG1)
    │   │   ├── cloudflare/              # Cloudflare Workers, Agents SDK
    │   │   ├── trailofbits/             # Trail of Bits security skills
    │   │   ├── qedgen/                # QEDGen formal verification (Lean 4)
    │   │   ├── solana-mobile/           # Mobile Wallet Adapter, Genesis Token
    │   │   ├── colosseum/              # Colosseum Copilot (startup research)
    │   │   ├── safe-solana-builder/   # Security-first code generation
    │   │   ├── vercel/                # Vercel deployment, Next.js, AI SDK
    │   │   ├── solana-new/            # SendAI idea→launch journey skills + datasets
    │   │   ├── ghostsecurity/         # Ghost Security AppSec skills
    │   │   ├── defending-code/        # Anthropic vuln-discovery reference harness
    │   │   ├── jupiter/               # Official Jupiter skills (swap, lend, VRFD)
    │   │   ├── metaplex/              # Official Metaplex (NFT, candy machine)
    │   │   ├── helius/                # Official Helius infra + SVM internals
    │   │   ├── quicknode-anchor/      # Anchor/Quasar reference files (quarantined)
    │   │   └── eth-to-sol/            # EVM/Solidity → Anchor porting
    │   ├── skill-registry.json     # Opt-in add-on catalog (repos/skills/MCPs)
    │   ├── idea-sprint/             # Wrapper: find + validate crypto ideas (GTM)
    │   ├── pitch-deck/              # Wrapper: pitch decks for crypto projects (GTM)
    │   ├── hackathon/               # Wrapper: hackathon submissions + grants (GTM)
    │   ├── token-2022.md            # Token Extensions guide (local)
    │   ├── backend-async.md         # Axum/Tokio patterns (local)
    │   └── deployment.md            # Deployment workflows (local)
    ├── rules/                   # Auto-loading constraints
    └── settings.json            # Permissions, hooks, agent teams

Agents

Commands

Building

Testing & Quality

Deployment & Migration

Workflow & Setup

DX Scripts

GitHub Action for Team Collaboration

This config includes a pre-built GitHub Action (.github/workflows/claude-code.yml) for PR-based iteration:

1. Add ANTHROPIC_API_KEY to your repository secrets
2. Copy .github/workflows/claude-code.yml to your project
3. Team members can @claude in PR comments
4. Claude responds with code suggestions using this configuration

Branch Workflow

All new work starts on a feature branch:

# Format: <type>/<scope>-<description>-<DD-MM-YYYY>
git checkout -b feat/program-vault-15-01-2026
git checkout -b fix/frontend-auth-15-01-2026

Use /quick-commit to automate branch creation and commits.

Code Quality

Before merging, run /diff-review or check diff against main:

git diff main...HEAD

Remove: excessive comments, abnormal try/catch blocks, verbose errors, redundant validation.

Keep: legitimate security checks, non-obvious explanations, matching error patterns.

Using as a GitHub Template

1. Click "Use this template" on GitHub (or fork the repo)
2. Clone your new repo
3. Run Claude Code and use /cleanup — this copies CLAUDE-solana.md → CLAUDE.md and removes config repo scaffolding (tests, install scripts, docs)
4. Start building!

For monorepos, add a CLAUDE.md to each package/module with architecture decisions scoped to that directory. Claude Code automatically loads these when working in that subdirectory.

Updating

# Preferred: use /update in Claude Code, or run directly:
bash .claude/bin/update.sh

# Preview changes without applying:
bash .claude/bin/update.sh --dry-run

# Resync skill submodules only:
bash .claude/bin/resync.sh

# Or manually update submodules
git submodule update --remote --merge

License

MIT - See LICENSE