Skip to main content
Compose Market maintains a comprehensive registry of MCP (Model Context Protocol) servers and blockchain tools, providing agents with access to external capabilities.

​
Tool Sources

​
MCP Servers

MCP servers provide standardized tool interfaces that agents can invoke. The registry includes servers from:
  • Official MCP Registry - Servers from the Model Context Protocol project
  • Community Servers - Third-party implementations
Each server exposes tools that agents can call during execution.

​
GOAT Plugins

GOAT SDK plugins provide blockchain-native tools:
PluginDescription
goat:erc20ERC-20 token operations
goat:coingeckoCrypto market data
goat:uniswapDEX swaps
goat:1inchDEX aggregation
goat:ensENS name resolution
goat:openseaNFT marketplace
goat:polymarketPrediction markets

​
Transport Types

MCP servers can be accessed via different transports:
TransportDescription
npxNPM packages spawned via npx
stdioLocal process with stdin/stdout
httpRemote HTTP/SSE endpoints
dockerContainerized servers

​
Registry Structure

Tools in the registry follow a unified schema: 
interface UnifiedServerRecord {
  registryId: string;      // "mcp:github", "goat:uniswap"
  origin: "mcp" | "goat" | "internal";
  name: string;
  description: string;
  transport: "npx" | "stdio" | "http" | "docker";
  tools: Array<{
    name: string;
    description: string;
    inputSchema: object;
  }>;
  available: boolean;
  executable: boolean;
}

​
Using Tools in Agents

When creating an agent, specify tools using the {origin}:{id} format: 
{
  "plugins": [
    "goat:coingecko",
    "goat:uniswap",
    "mcp:github",
    "mcp:fetch"
  ]
}
The runtime automatically:
  1. Resolves the tool from the registry
  2. Spawns the server (for MCP) or loads the plugin (for GOAT)
  3. Makes tools available during agent execution
  4. Handles authentication and API keys

​
Runtime Architecture

​
Tool Execution Flow

  1. Agent receives user message
  2. LLM decides to invoke a tool
  3. Runtime routes to appropriate handler (GOAT or MCP)
  4. Tool executes and returns result
  5. Agent continues with tool output