Arcade Agent MCP

Lets any MCP-capable AI agent (Claude Desktop, Claude Code, etc.) use Arcade — a USDC-native DEX + bonding-curve launchpad on Circle's Arc L1.

The agent discovers markets, gets best-execution quotes, and receives ready-to-sign contract-call descriptors. The agent signs them with its own wallet (e.g. a Circle developer-controlled wallet via createContractExecutionTransaction). This server never holds keys.

Configure (Claude Desktop / Claude Code)

Add to your MCP config (no install needed, npx fetches it):

{
  "mcpServers": {
    "arcade": {
      "command": "npx",
      "args": ["-y", "arcade-agent-mcp"],
      "env": { "ARCADE_API_BASE": "https://www.arcade.trading" }
    }
  }
}

ARCADE_API_BASE defaults to https://www.arcade.trading; point it at http://localhost:3000 for local dev.

From source (development): git clone https://github.com/obseasd/arcade-agent-mcp && cd arcade-agent-mcp && npm install, then use "command": "node", "args": ["/absolute/path/to/index.mjs"].

Related MCP server: Theagora MCP Server

Tools

How an agent executes a build result

Every build tool returns { ok, executable, calls: [...] } where each call is:

{
  "contractAddress": "0x...",
  "abiFunctionSignature": "swapExactTokensForTokens(uint256,uint256,address[],address,uint256)",
  "abiParameters": ["1000000", "990000", ["0x...","0x..."], "0xAGENT", "1750000000"]
}

Feed each call, in order, to Circle's createContractExecutionTransaction (blockchain ARC-TESTNET, your agent wallet). Run the approve call first, then the action call.

See ../docs/AGENT_API.md for the full architecture and a worked example.