Agent Skills

What Are Skills?

Skills are pre-packaged, multi-step workflows that AI agents can follow to accomplish complex DeFi tasks on Starknet. Each skill contains a SKILL.md file with step-by-step instructions, prerequisites, error handling, and expected outcomes.

Unlike MCP tools (which are individual function calls), skills orchestrate multiple tools and decisions into a complete workflow — for example, checking auth status → fetching a swap quote → confirming with the user → executing → verifying the transaction.

Installation

Install all 10 StarkFi skills into your AI assistant with a single command:

npx skills add ahmetenesdur/starkfi

This uses Vercel's Skills CLI to add the skills directly to your agent (e.g., Antigravity, Claude Code).

Prerequisites

Node.js v18 or later
StarkFi Auth Server running (required for authentication and transaction signing)
No manual CLI installation needed — all skills use npx starkfi@latest internally

Available Skills

Skill	Category	Description
authenticate-wallet	auth	Email OTP login, session management, wallet deploy
balance	wallet-data	Check STRK, ETH, and ERC-20 token balances
send	transaction	Transfer tokens to a Starknet address
trade	transaction	Swap tokens via Fibrous aggregation
multi-swap	transaction	Multiple swaps in one transaction (up to 3 pairs)
batch	transaction	Combine swap + stake + supply + send in one multicall
staking	transaction	Stake, unstake, claim, compound rewards
lending	transaction	Vesu V2: supply, borrow, repay, withdraw, close
portfolio	wallet-data	Aggregated DeFi dashboard with USD valuations
config	utility	RPC, network, gas settings, transaction status

Getting Started

Typical Workflow

Skills follow a natural progression — authenticate first, then check your state, then execute operations:

1. authenticate-wallet  →  Login and verify your session
2. balance / portfolio  →  Check funds and existing positions
3. trade / send / ...   →  Execute DeFi operations
4. config (tx-status)   →  Verify transaction results

Trigger Examples

Your AI agent automatically selects the right skill based on what you say:

You say...	Skill triggered
"Log me in with my email"	`authenticate-wallet`
"How much STRK do I have?"	`balance`
"Send 10 USDC to 0x04a3..."	`send`
"Swap 100 USDC for ETH"	`trade`
"Swap USDC→ETH and USDT→STRK"	`multi-swap`
"Swap ETH then stake STRK"	`batch`
"Stake 1000 STRK with Karnot"	`staking`
"Supply 500 USDC to Prime pool"	`lending`
"Show me my portfolio"	`portfolio`
"I'm getting rate limit errors"	`config`

Skill Structure

Each skill is a self-contained directory with a SKILL.md file:

skills/
├── authenticate-wallet/
│   └── SKILL.md
├── trade/
│   └── SKILL.md
├── staking/
│   └── SKILL.md
└── ...

Every SKILL.md follows a consistent structure:

Section	Purpose
YAML Frontmatter	Name, description (triggers AI matching), and metadata
Prerequisites	Required auth state, minimum versions
Rules	Hard constraints the agent must follow
Commands	CLI commands with full syntax
Parameters	Options table with types and defaults
Examples	Real-world command examples
Error Handling	Error → cause → resolution table
Related Skills	Cross-references to complementary skills

How Agents Use Skills

When an MCP-compatible agent (like Claude, Cursor, or Windsurf) connects to StarkFi:

The agent discovers available skills via the skills/ directory
It matches the user's intent to the best skill using the description field
The SKILL.md provides a structured recipe — the agent follows it step by step
Each step maps to a CLI command (via npx starkfi@latest)
Error handling tables tell the agent exactly what to do when things fail

Skills vs MCP Tools

	Skills	MCP Tools
Scope	Multi-step workflows	Single function calls
Format	Markdown instructions	JSON-schema typed tools
Usage	Agent reads and follows	Agent calls directly
Example	"Complete a swap" (quote → confirm → execute → verify)	`swap_tokens(args)`
Best for	Complex workflows with decisions	Simple, atomic operations

Skills and MCP tools are complementary — an agent can use skills for guided workflows and MCP tools for ad-hoc operations.

On this page