mcpc: Instructions file for GitHub Copilot

MCPC Development Guide

MCPC lets you build agentic MCP servers by composing existing MCP tools. Think of it as creating AI agents from reusable tool libraries, the agent itself is a MCP tool too.

Rules

NEVER use deno specific APIs in code, use node compatible APIs only.

Core Pattern

Every MCPC server follows the same simple pattern:

const server = await mcpc(
  [{ name: "my-agent", version: "1.0.0" }, {
    capabilities: { tools: {} },
  }],
  [{
    name: "my-agent",
    description: "Agent docs with <tool name='server.tool'/> references",
    deps: { mcpServers: {/* MCP server configs */} },
    options: { mode: "agentic" },
  }],
);
await server.connect(new StdioServerTransport());

That's it. Three parts: server metadata, agent definition(s), transport connection.

Tool Selection with XML Tags

Reference tools in your agent description using XML-like syntax:

description: `
Available tools:
<tool name="desktop-commander.read_file"/>
<tool name="github.__ALL__"/>  // All tools from github server
<tool name="tool1" maxResultLength="2000"/>  // Limit result size
<tool name="sensitive" hide/>  // Hide from public interface
`;

Tool names are always servername.toolname. The framework auto-maps both server.tool and server_tool formats.

MCP Server Dependencies

Support all transport types: stdio, streamable-http, sse.

deps: {
  mcpServers: {
    "desktop-commander": {
      command: "npx",
      args: ["-y", "@wonderwhy-er/desktop-commander@latest"],
      transportType: "stdio"
    },
    "github": {
      transportType: "streamable-http",
      url: "https://api.githubcopilot.com/mcp/",
      headers: { "Authorization": "Bearer ${GITHUB_TOKEN}" }  // Env vars work
    }
  }
}

Execution Modes

• mode: "agentic" (default): Interactive tool calls, agent decides next step
• mode: "agentic_workflow": Structured steps (predefined or generated at runtime)
• sampling: true: Autonomous execution in compatible clients (VS Code, etc.)