Claude Agent SDK vs OpenAI Agents SDK Comparison

// Claude Agent SDK (TypeScript) - File-based subagent + MCP tool restriction
// Source: docs.claude.com/en/api/agent-sdk/subagents + /mcp
import { query } from "@anthropic-ai/claude-agent-sdk";

const result = query({
  prompt: "Find and fix the failing tests in the repo",
  options: {
    // Programmatic subagent definition — file-based alternative: .claude/agents/*.md
    agents: {
      "test-fixer": {
        description: "Analyzes and fixes failing tests",
        prompt: "You are a test engineer. Run the test first, then find the root cause.",
        tools: ["Read", "Edit", "Bash"], // tool restriction
      },
    },
    // MCP server: connect as a local process
    mcpServers: {
      github: {
        command: "npx",
        args: ["-y", "@modelcontextprotocol/server-github"],
      },
    },
    permissionMode: "default", // hooks + deny/ask/allow order kicks in
  },
});

for await (const message of result) {
  if (message.type === "result") {
    console.log("Total cost:", message.total_cost_usd);
    console.log("Model usage:", message.modelUsage);
  }
}

# OpenAI Agents SDK (Python) - Handoff + guardrail + MCP (stdio)
# Source: openai.github.io/openai-agents-python/multi_agent/ + /mcp/ + /guardrails/
from agents import Agent, Runner, handoff, input_guardrail, GuardrailFunctionOutput
from agents.mcp import MCPServerStdio

async def main():
    async with MCPServerStdio(
        params={"command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"]},
    ) as github_mcp:

        @input_guardrail
        async def block_secrets(ctx, agent, input_text: str) -> GuardrailFunctionOutput:
            contains_secret = "sk-" in input_text
            return GuardrailFunctionOutput(
                output_info={"blocked": contains_secret},
                tripwire_triggered=contains_secret,
            )

        triage_agent = Agent(
            name="Triage",
            instructions="Analyze the task and hand off to the right specialist agent.",
            input_guardrails=[block_secrets],
        )

        fixer_agent = Agent(
            name="TestFixer",
            instructions="Fix the failing tests.",
            mcp_servers=[github_mcp],
        )

        triage_agent.handoffs = [handoff(fixer_agent)]

        result = await Runner.run(triage_agent, "Fix the failing tests in the repo")
        print(result.final_output)