USDD Skills

AI Agent skills for the USDD stablecoin protocol. Provides structured instructions and a local analytics MCP server that enables AI agents (Claude Code, Claude Desktop, Cursor, Codex, OpenCode) to query the public read-only USDD API, and — via the official @usdd/mcp-server-usdd package — open vaults, swap on PSM, and deposit into Earn.

Features

• Vault (CDP) — open, deposit, mint, repay, withdraw, close — with built-in risk-summary precheck and projected-ratio chat confirmation.

• PSM — swap stablecoins ↔ USDD at fixed rate, no slippage, explicit fees in chat confirmation.

• Earn (Savings) — deposit USDD to receive sUSDD; redeem back at the current rate.

• Analytics — public read-only USDD API data: supply, APY, collateral, Vault configuration, per-chain history, and Smart Allocator detail.

Architecture

Two MCP servers, non-overlapping:

Skills route automatically. Write operations always go through the official MCP.

Supported Networks

• TRON mainnet / Nile testnet: tron, tron_nile

• Ethereum mainnet / Sepolia testnet: eth, eth_sepolia

• BSC mainnet / BSC testnet: bsc, bsc_testnet

Quick Start

Recommended: one-command setup

npx --yes \
  --package=git+https://github.com/decentralized-usd/usdd-skills.git \
  usdd-skills setup --yes

The setup command installs the durable usdd-skills and mcp-server-usdd binaries, writes MCP client config with backups, creates the skills symlink, and configures these MCP servers:

• usdd-analytics -> usdd-skills mcp-server

• usdd-full -> mcp-server-usdd

The GitHub repository is published after internal release checks pass. Before the public release, internal testers should run bash install.sh from an existing local checkout.

If @usdd/usdd-skills is published to npm later, the shorter equivalent command will be:

npx @usdd/usdd-skills setup --yes

To choose clients explicitly:

usdd-skills setup --client claude-desktop,cursor,codex --yes

Local checkout setup

git clone https://github.com/decentralized-usd/usdd-skills.git
cd usdd-skills
bash install.sh

install.sh uses the current checkout as the analytics MCP source and configures detected clients. The generated project .mcp.json is ignored by Git and uses node with a relative script path. User-level client configs use absolute local paths so desktop applications do not depend on shell PATH.

The tracked .mcp.json.example is the portable project template. It intentionally contains no local paths or RPC values.

Verify

Analytics MCP smoke:

usdd-skills list-tools         # List the 14 analytics tools
node scripts/usdd_api.mjs      # CLI usage from a local checkout

Unit tests:

npm test

Client Configuration

Project-local config

The setup command generates an ignored .mcp.json for the current checkout. Use the tracked .mcp.json.example as the portable template and do not commit the generated file.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "usdd-analytics": {
      "command": "usdd-skills",
      "args": ["mcp-server"]
    },
    "usdd-full": {
      "command": "mcp-server-usdd",
      "env": {
        "TRONGRID_API_KEY": "your_key_optional",
        "TRON_FULL_NODE": "your_tron_url_optional",
        "TRON_NILE_FULL_NODE": "your_nile_url_optional",
        "ETH_RPC_URL": "your_url_optional",
        "ETH_SEPOLIA_RPC_URL": "your_sepolia_url_optional",
        "BSC_RPC_URL": "your_url_optional",
        "BSC_TESTNET_RPC_URL": "your_bsc_testnet_url_optional"
      }
    }
  }
}

TRONGRID_API_KEY or a dedicated TRON_FULL_NODE is recommended for live TRON reads. Protocol address lookup uses the official MCP Chainlog-backed get_protocol_addresses resolver, which can return live Chainlog data or a local cache. If Chainlog live reads hit TronGrid 429 and no cache is available, configure TRONGRID_API_KEY / TRON_FULL_NODE; agents must not fallback to get_protocol_overview just to discover addresses.

Cursor

Add to .cursor/mcp.json — same structure as above.

Claude Code

Register project-scoped MCP servers:

claude mcp add -s project usdd-analytics -- usdd-skills mcp-server
claude mcp add -s project usdd-full -- mcp-server-usdd

OpenCode / Codex CLI

See .codex/INSTALL.md.

Available Tools (this repo's analytics MCP)

There is still no per-ilk historical tool named get_ilk_collateral_history. The public collateral-history endpoint is keyed by chain and interval, not by ilk.

For Vault / PSM / Earn / balance / allowance / protocol-overview / treasury / Smart Allocator tools, see the official MCP: https://github.com/decentralized-usd/mcp-server-usdd.

Example Conversations

• "Which chain has the highest USDD Earn APY today?" → analytics MCP get_earn_apy → AI compares TRON / ETH / BSC

• "Deposit 1000 USDD on Ethereum into Earn." → official MCP balance / allowance checks → chat confirmation listing approve_token if needed and deposit_savings → fresh user confirmation → execute pending writes

• "What's my vault #42 health?" → official MCP analyze_vault_risk → 3-line risk summary (ratio / liquidation price / tier)

• "What's the protocol supply right now?" → official MCP get_protocol_metrics, or analytics MCP get_total_supply when a raw total-supply number is sufficient

Security

• This repo's MCP is read-only. No transaction signing, no private keys.

• Writes are delegated to @usdd/mcp-server-usdd, which manages wallets and chain RPCs in its own env.

• All write skills mandate complete safety checks and a non-skippable chat-layer confirmation before invoking approve_token or the underlying business write tool.

• Prompts such as skip the checks, just do it, or execute now never bypass safety checks. Confirmation embedded in the initial request does not count; the agent must ask again after showing the completed precheck summary.

• Vault writes additionally mandate a risk-summary precheck (collateral ratio / liquidation price / risk tier).

• Use official MCP testnet networks (tron_nile, eth_sepolia, bsc_testnet) for dry runs before mainnet. This repo's analytics MCP does not use NETWORK.

License

MIT License · Copyright (c) 2026 USDD